Scientific frontiers of agentic AI

The language AI agents might speak, sharing context without compromising privacy, modeling agentic negotiations, and understanding users’ commonsense policies are some of the open scientific questions that researchers in agentic AI will need to grapple with.

It feels as though we’ve barely absorbed the rapid development and adoption of generative AI technologies such as large language models (LLMs) before the next phenomenon is already upon us, namely agentic AI. Standalone LLMs can be thought of as “chatbots in a sandbox”, the sandbox being a metaphor for a safe and contained play space with limited interaction with the world beyond. In contrast, the vision of agentic AI is a near (or already here?) future in which LLMs are the underlying engines for complex systems that have access to rich external resources such as consumer apps and services, social media, banking and payment systems — in principle, anything you can reach on the Internet. A dream of the AI industry for decades, the “agent” of agentic AI is an intelligent personal assistant that knows your goals and preferences and that you trust to act on your behalf in the real world, much as you might a human assistant.

Related content
Real-world deployment requires notions of fairness that are task relevant and responsive to the available data, recognition of unforeseen variation in the “last mile” of AI delivery, and collaboration with AI activists.

For example, in service of arranging travel plans, my personal agentic AI assistant would know my preferences (both professional and recreational) for flights and airlines, lodging, car rentals, dining, and activities. It would know my calendar and thus be able to schedule around other commitments. It would know my frequent-flier numbers and hospitality accounts and be able to book and pay for itineraries on my behalf. Most importantly, it would not simply automate these tasks but do so intelligently and intuitively, making “obvious” decisions unilaterally and quietly but being sure to check in with me whenever ambiguity or nuance arises (such as whether those theater tickets on a business trip to New York should be charged to my personal or work credit card).

To AI insiders, the progression from generative to agentic AI is exciting but also natural. In just a few years, we have gone from impressive but glorified chatbots with myriad identifiable shortcomings to feature-rich systems exhibiting human-like capabilities not only in language and image generation but in coding, mathematical reasoning, optimization, workflow planning, and many other areas. The increased skill set and reliability of core LLMs has naturally caused the industry to move “up the stack”, to a world in which the LLM itself fades into the background and becomes a new kind of intelligent operating system upon which all manner of powerful functionality can be built. In the same way that your PC or Mac seamlessly handles many details that the vast majority of users don’t (want to) know about — like exactly how and where on your hard drive to store and find files, the networking details of connecting to remote web servers, and other fine-grained operating-system details — agentic systems strive to abstract away the messy and tedious details of many higher-level tasks that, today, we all perform ourselves.

But while the overarching vision of agentic AI is already relatively clear, there are some fundamental scientific and technical questions about the technology whose answers — or even proper formulation — are uncertain (but interesting!). We’ll explore some of them here.

What language will agents speak?

The history of computing technology features a steady march toward systems and devices that are ever more friendly, accessible, and intuitive to human users. Examples include the gradual displacement of clunky teletype monitors and obscure command-line incantations by graphical user interfaces with desktop and folder metaphors, and the evolution from low-level networked file transfer protocols to the seamless ease of the web. And generative AI itself has also made previously specialized tasks like coding accessible to a much broader base of users. In other words, modern technology is human-centric, designed for use and consumption by ordinary people with little or no specialized training.

But now these same technologies and systems will also need to be navigated by agentic AI, and as adept as LLMs are with human language, it may not be their most natural mode of communication and understanding. Thus, a parallel migration to the native language of generative AI may be coming.

What is that native language? When generative AI consumes a piece of content — whether it be a user prompt, a document, or an image — it translates it into an internal representation that is more convenient for subsequent processing and manipulation. There are many examples in biology of such internal representations. For instance, in our own visual systems, it has been known for some time that certain types of inputs (such as facial images) cause specific cells in our brains to respond (a phenomenon known as neuronal selectivity). Thus, an entire category of important images elicits similar neural behaviors.

Related content
Generative AI raises new challenges in defining, measuring, and mitigating concerns about fairness, toxicity, and intellectual property, among other things. But work has started on the solutions.

In a similar vein, the neural networks underlying modern AI typically translate any input into what is known as an embedding space, which can be thought of as a physical map in which items with similar meanings are placed near each other, and those with unrelated meanings are placed far apart. For example, in an image-embedding space, two photos of different families would be nearer to each other than either would be to a landscape. In a language-embedding space, two romance novels would be nearer to each other than to a car owner’s manual. And hybrid or multimodal embedding spaces would place images of cars near their owner manuals.

Embeddings are an abstraction that provides great power and generality, in the form of the ability to represent not the literal original content (like a long sequence of words) but something closer to its underlying meaning. The price for this abstraction is loss of detail and information. For instance, the embedding of this entire article would place it in close proximity to similar content (for instance, general-audience science prose) but would not contain enough information to re-create the article verbatim. The lossy nature of embeddings has implications we shall return to shortly.

Embeddings are learned from the massive amount of information on the Internet and elsewhere about implicit correspondences. Even aliens landing on earth who could read English but knew nothing else about the world would quickly realize that “doctor” and “hospital” are closely related because of their frequent proximity in text, even if they had no idea what these words actually signified. Furthermore, not only do embeddings permit generative AI to understand existing content, but they allow it to generate new content. When we ask for a picture of a squirrel on a snowboard in the style of Andy Warhol, it is the embedding that lets the technology explore novel images that interpolate between those of actual Warhols, squirrels, and snowboards.

Thus, the inherent language of generative (and therefore agentic) AI is not the sentences and images we are so familiar with but their embeddings. Let us now reconsider a world in which agents interact with humans, content, and other agents. Obviously, we will continue to expect agentic AI to communicate with humans in ordinary language and images. But there is no reason for agent-to-agent communication to take place in human languages; per the discussion above, it would be more natural for it to occur in the native embedding language of the underlying neural networks.

My personal agent, working on a vacation itinerary, might ingest materials such as my previous flights, hotels, and vacation photos to understand my interests and preferences. But to communicate those preferences to another agent — say, an agent aggregating hotel details, prices, and availability — it will not provide the raw source materials; in addition to being massively inefficient and redundant, that could present privacy concerns (more on this below). Rather, my agent will summarize my preferences as a point, or perhaps many points, in an embedding space.

Restaurant embeddings.jpg
In this example, the red, green, and blue points are three-dimensional embeddings of restaurants at which three people (Alice, Bob, and Chris) have eaten. (A real-world embedding, by contrast, might have hundreds of dimensions.) Each glowing point represents the center of one of the clusters, and its values summarize the restaurant preferences of the corresponding person. AI agents could use such vector representations, rather than text, to share information with each other.

By similar reasoning, we might also expect the gradual development of an “agentic Web” meant for navigation by AI, in which the text and images on websites are pre-translated into embeddings that are illegible to humans but are massively more efficient than requiring agents to perform these translations themselves with every visit. In the same way that many websites today have options for English, Spanish, Chinese, and many other languages, there would be an option for Agentic.

All the above presupposes that embedding spaces are shared and standardized across generative and agentic AI systems. This is not true today: embeddings differ from model to model and are often considered proprietary. It’s as if all generative AI systems speak slightly different dialects of some underlying lingua franca. But these observations about agentic language and communication may foreshadow the need for AI scientists to work toward standardization, at least in some form. Each agent can have some special and proprietary details to its embeddings — for instance, a financial-services agent might want to use more of its embedding space for financial terminology than an agentic travel assistant would — but the benefits of a common base embedding are compelling.

Keeping things in context

Even casual users of LLMs may be aware of the notion of “context”, which is informally what and how much the LLM remembers and understands about its recent interactions and is typically measured (at least cosmetically) by the number of words or tokens (word parts) recalled. There is again an apt metaphor with human cognition, in the sense that context can be thought of as the “working memory” of the LLM. And like our own working memory, it can be selective and imperfect.

If we participate in an experiment to test how many random digits or words we can memorize at different time scales, we will of course eventually make mistakes if asked to remember too many things for too long. But we will not forget what the task itself is; our short-term memory may be fallible, but we generally grasp the bigger picture.

Related content
Large language models’ emergent abilities are improving with scale; as scale grows, where are LLMs heading? Insights from Ray Solomonoff’s theory of induction and stochastic realization theory may help us envision — and guide — the limits of scaling.

These same properties broadly hold for LLM context — which is sometimes surprising to users, since we expect computers to be perfect at memorization but highly fallible on more abstract tasks. But when we remember that LLMs do not operate directly on the sequence of words or tokens in the context but on the lossy embedding of that sequence, these properties become less mysterious (though perhaps not less frustrating when an LLM can’t remember something it did just a few steps ago).

Some of the principal advances in LLM technology have been around improvements in context: LLMs can now remember and understand more context and leverage that context to tailor their responses with greater accuracy and sophistication. This greater window of working memory is crucial for many tasks to which we would like to apply agentic AI, such as having an LLM read and understand the entire code base of a large software development project, or all the documents relevant to a complex legal case, and then be able to reason about the contents.

How will context and its limitations affect agentic AI? If embeddings are the language of LLMs, and context is the expression of an LLM’s working memory in that language, a crucial design decision in agent-agent interactions will be how much context to share. Sharing too little will handicap the functionality and efficiency of agentic dialogues; sharing too much will result in unnecessary complexity and potential privacy concerns (just as in human-to-human interactions).

Let us illustrate by returning to my personal agent, who having found and booked my hotel is working with an external airline flight aggregation agent. It would be natural for my agent to communicate lots of context about my travel preferences, perhaps including conditions under which I might be willing to pay or use miles for an upgrade to business class (such as an overnight international flight). But my agent should not communicate context about my broader financial status (savings, debt, investment portfolio), even though in theory these details might correlate with my willingness to pay for an upgrade. When we consider that context is not my verbatim history with my travel agent, but an abstract summary in embedding space, decisions about contextual boundaries and how to enforce them become difficult.

Indeed, this is a relatively untouched scientific topic, and researchers are only just beginning to consider questions such as what can be reverse-engineered about raw data given only its embedding. While human or system prompts to shape inter-agent dealings might be a stopgap (“be sure not to tell the flight agent any unnecessary financial information”), a principled understanding of embedding privacy vulnerabilities and how to mitigate them (perhaps via techniques such as differential privacy) is likely to be an important research area going forward.

Agentic bargains

So far, we’ve talked a fair amount about interagent dialogues but have treated these conversations rather generally, much as if we were speaking about two humans in a collaborative setting. But there will be important categories of interaction that will need to be more structured and formal, with identifiable outcomes that all parties commit to. Negotiation, bargaining, and other strategic interactions are a prime example.

I obviously want my personal agent, when booking hotels and flights for my trips, to get the best possible prices and other conditions (room type and view, flight seat location, and so on). The agents aggregating hotels and flights would similarly prefer that I pay more rather than less, on behalf of their own clients and users.

For my agent to act in my interests in these settings, I’ll need to specify at least some broad constraints on my preferences and willingness to pay for them, and not in fuzzy terms: I can’t expect my agent to simply “know a bargain when it sees one” the way I might if I were handling all the arrangements myself, especially because my notion of a bargain might be highly subjective and dependent on many factors. Again, a near-term makeshift approach might address this via prompt shaping — “be sure to get the best deal possible, as long as the flight is nonstop and leaves in the morning, and I have an aisle seat” — but longer-term solutions will have to be more sophisticated and granular.

Related content
Amazon Research Award recipient Éva Tardos studies complex theoretical questions that have far-ranging practical consequences.

Of course, the mathematical and scientific foundations of negotiating and bargaining have been well studied for decades by game theorists, microeconomists, and related research communities. Their analyses typically begin by presuming the articulation of utility functions for all the parties involved — an abstraction capturing (for example) my travel preferences and willingness to pay for them. The literature also considers settings in which I can’t quantitatively express my own utilities but “know bargains when I see them”, in the sense that given two options (a middle seat on a long flight for $200 vs. a first-class seat for $2,000), I will make the choice consistent with my unknown utilities. (This is the domain of the aptly named utility elicitation.)

Much of the science in such areas is devoted to the question of what “should” happen when fully rational parties with precisely specified utilities, perfect memory, and unlimited computational power come to the proverbial bargaining table; equilibrium analysis in game theory is just one example of this kind of research. But given our observations about the human-like cognitive abilities and shortcomings of LLMs, perhaps a more relevant starting point for agentic negotiation is the field of behavioral economics. Instead of asking what should happen when perfectly rational agents interact, behavioral economics asks what does happen when actual human agents interact strategically. And this is often quite different, in interesting ways, than what fully rational agents would do.

For instance, consider the canonical example of behavioral game theory known as the ultimatum game. In this game, there is $10 to potentially divide between two players, Alice and Bob. Alice first proposes any split she likes. Bob then either accepts Alice’s proposal, in which case both parties get their proposed shares, or rejects Alice’s proposal, in which case each party receives nothing. The equilibrium analysis is straightforward: Alice, being fully rational and knowing that Bob is also, proposes the smallest nonzero amount to Bob, which is a penny. Bob, being fully rational, would prefer to receive a penny than nothing, so he accepts.

Ultimatum game 1.jpg
Game theory (left) supposes that the recipient in the ultimatum game will accept a low offer, since something is better than nothing, but behavioral economics (right) reveals that, in fact, offers tend to concentrate in the range of $3 to $5, and lower offers are frequently rejected.

Nothing remotely like this happens when humans play. Across hundreds of experiments varying myriad conditions — social, cultural, gender, wealth, etc. — a remarkably consistent aggregate behavior emerges. Alice almost always proposes a share to Bob of between $3 and $5 (the fact that Alice gets to move first seems to prime both players for Bob to potentially get less than half the pie). And conditioned on Alice’s proposal being in this range, Bob almost always accepts her offer. But on those rare occasions in which Alice is more aggressive and offers Bob an amount much less than $3, Bob’s rejection rate skyrockets. It’s as if pairs of people — who have never heard of or played the ultimatum game before — have an evolutionarily hardwired sense of what’s “fair” in this setting.

Ultimatum game bar graph.jpg
The way in which the ultimatum game is played — the frequency of particular offers and the rate of rejection — varies across cultures, but this graph illustrates general trends in the data. Offers tend to concentrate between $3 and $5, with a steep falloff above $5, and the rejection rate is high for low offers.

Now back to LLMs and agentic AI. There is already a small but growing literature on what we might call LLM behavioral game theory and economics, in which experiments like the one above are replicated — except human participants are replaced by AI. One early work showed that LLMs almost exactly replicated human behavior in the ultimatum game, as well as other classical behavioral-economics findings.

Note that it is possible to simulate the demographic variability of human subjects in such experiments via LLM prompting, e.g., “You are Alice, a 37-year-old Hispanic medical technician living in Boston, Massachusetts”. Other studies have again shown human-like behavior of LLMs in trading games, price negotiations, and other settings. A very recent study claims that LLMs can even engage in collusive price-fixing behaviors and discusses potential regulatory implications for AI agents.

Once we have a grasp on the behaviors of agentic AI in strategic settings, we can turn to shaping that behavior in desired ways. The field of mechanism design in economics complements areas like game theory by asking questions like “given that this is how agents generally negotiate, how can we structure those negotiations to make them fair and beneficial?” A classic example is the so-called second-price auction, where the highest bidder wins the item — but only pays the second highest bid. This design is more truthful than a standard first-price auction, in the sense that everyone’s optimal strategy is to simply bid the price at which they are indifferent to winning or losing (their subjective valuation of the item); nobody needs to think about other agents’ behaviors or valuations.

We anticipate a proliferation of research on topics like these, as agentic bargaining becomes commonplace and an important component of what we delegate to our AI assistants.

The enduring challenge of common sense

I’ll close with some thoughts on a topic that has bedeviled AI from its earliest days and will continue to do so in the agentic era, albeit in new and more personalized ways. It’s a topic that is as fundamental as it is hard to define: common sense.

By common sense, we mean things that are “obvious”, that any human with enough experience in the world would know without explicitly being told. For example, imagine a glass full of water sitting on a table. We would all agree that if we move the glass to the left or right on the table, it’s still a glass of water. But if we turn it upside down, it’s still a glass on the table, but no longer a glass of water (and is also a mess to be cleaned up). It’s quite unlikely any of us were ever sat down and run through this narrative, and it’s also a good bet that you’ve never deliberately considered such facts before. But we all know and agree on them.

Related content
Using large language models to discern commonsense relationships can improve performance on downstream tasks by as much as 60%.

Figuring out how to imbue AI models and systems with common sense has been a priority of AI research for decades. Before the advent of modern large-scale machine learning, there were efforts like the Cyc project (for “encyclopedia”), part of which was devoted to manually constructing a database of commonsense facts like the ones above about glasses, tables, and water. Eventually the consumer Internet generated enough language and visual data that many such general commonsense facts could be learned or inferred: show a neural network millions of pictures of glasses, tables and water and it will figure things out. Very early research also demonstrated that it was possible to directly encode certain invariances (similar to shifting a glass of water on a table) into the network architecture, and LLM architectures are similarly carefully designed in the modern era.

But in agentic AI, we expect our proxies to understand not only generic commonsense facts of the type we’ve been discussing but also “common sense” particular to our own preferences — things that would make sense to most people if only they understood our contexts and perspectives. Here a pure machine learning approach will likely not suffice. There just won’t be enough data to learn from scratch my subjective version of common sense.

For example, consider your own behavior or “policy” around leaving doors open or closed, locked or unlocked. If you’re like me, these policies can be surprisingly nuanced, even though I follow them without thought all the time. Often, I will close and lock doors behind me — for instance, when I leave my car or my house (unless I’m just stepping right outside to water the plants). Other times I will leave a door unlocked and open, such as when I’m in my office and want to signal I am available to chat with colleagues or students. I might close but leave unlocked that same door when I need to focus on something or take a call. And sometimes I’ll leave my office door unlocked and open even when I’m not in it, despite there being valuables present, because I trust the people on my floor and I’m going to be nearby.

We might call behaviors like these subjective common sense, because to me they are natural and obvious and have good reasons behind them, even though I follow them almost instinctually, the same way I know not to turn a glass of water upside down on the table. But you of course might have very different behaviors or policies in the same or similar situations, with your own good reasons.

Related content
Dataset contains more than 11,000 newly collected dialogues to aid research in open-domain conversation.

The point is that even an apparently simple matter like my behavior regarding doors and locks can be difficult to articulate. But agentic AI will need specifications like this: simply replace doors with online accounts and services and locks with passwords and other authentication credentials. Sometimes we might share passwords with family or friends for less-critical privacy-sensitive resources like Netflix or Spotify, but we would not do the same for bank accounts and medical records. I might be less rigorous about restricting access to, or even encrypting, the files on my laptop than I would be about files I store in the cloud.

The circumstances under which I trust my own or other agents with resources that need to be private and secure will be at least as complex as those regarding door closing and locking. The primary difficulty is not in having the right language or formalisms to specify such policies: there are good proposals for such specification frameworks and even for proving the correctness of their behaviors. The problem is in helping people articulate and translate their subjective common sense into these frameworks in the first place.

Conclusion

The agentic-AI era is in its infancy, but we should not take that to mean we have a long and slow development and adoption period before us. We need only look at the trajectory of the underlying generative AI technology — from being almost entirely unknown outside of research circles as recently as early 2022 to now being arguably the single most important scientific innovation of the century so far. And indeed, there is already widespread use of what we might consider early agentic systems, such as the latest coding agents.

Far beyond the initial “autocomplete for Python” tools of a few years ago, such agents now do so much more — writing working code from natural-language prompts and descriptions, accessing external resources and datasets, proactively designing experiments and visualizing the results, and most importantly (especially for a novice programmer like me), seamlessly handling the endless complexity of environment settings, software package installs and dependencies, and the like. My Amazon Scholar and University of Pennsylvania colleague Aaron Roth and I recently wrote a machine learning paper of almost 50 pages — complete with detailed definitions, theorem statements and proofs, code, and experiments — using nothing except (sometimes detailed) English prompts to such a tool, along with expository text we wrote directly. This would have been unthinkable just a year ago.

Despite the speed with which generative AI has permeated industry and society at large, its scientific underpinnings go back many decades, arguably to the birth of AI but certainly no later than the development of neural-network theory and practice in the 1980s. Agentic AI — built on top of these generative foundations, but quite distinct in its ambitions and challenges — has no such deep scientific substrate on which to systematically build. It’s all quite fresh territory. I’ve tried to anticipate some of the more fundamental challenges here, and I’ve probably got half of them wrong. To paraphrase the Philadelphia department store magnate John Wanamaker, I just don’t know which half — yet.

Related content

US, WA, Seattle
Amazon's eCommerce Foundation (eCF) organization provides the core technologies that drive and power Amazon's Stores, Digital, and Other (SDO) businesses. Millions of customer page views and orders per day are enabled by the systems eCF builds from the ground up. CloudTune, within eCF, empowers growth and business agility needs by automatically and efficiently managing AWS capacity and business processes needed to safely meet Amazon's customer demand. CloudTune serves its primary customers, internal software teams, through forecast-driven automation of cost controllership, capacity management, and scaling. We predict expected load, and drive procurement and allocation of AWS capacity for new product launches and high-velocity events like Prime Day and Cyber Monday. CloudTune is looking for a Data Scientist to join our forecasting and optimization team. The team develops sophisticated algorithms that combine machine learning-based demand forecasting with mathematical optimization to solve large-scale capacity planning problems under uncertainty. We work with massive datasets — spanning thousands of availability zone and instance family combinations across global regions — to determine optimal resource allocation strategies that balance infrastructure holding costs against service availability requirements. These models directly inform multi-million dollar capacity investment decisions and drive automated procurement, retention, and release policies across Amazon's compute infrastructure. As a Data Scientist in CloudTune, you will work with other scientists, software engineers, data engineers, and product managers on a variety of important research problems in the areas of stochastic optimization, time series modeling, and operations research. You will work on statistical problems with a high level of ambiguity. You will analyze and process large amounts of data, develop new algorithms and improve existing approaches based on statistical models, machine learning algorithms and big data solutions to automatically scale Amazon’s compute infrastructure, optimizing the balance between availability risk and cost efficiency for all of Amazon businesses. Key job responsibilities - Process and analyze large data sets, mining additional data sources as needed - Build statistical models and drive scalable solutions for multi-year capacity demand forecasting horizons - Create, enhance, and maintain technical documentation, and present to other scientists and business leaders - Process and analyze large-scale operational data to identify patterns and inform model development - Create, enhance, and maintain technical documentation, and present research findings to scientists, engineering teams, and senior leadership
US, CA, Irvine
Do you want to write the algorithms that control next-generation batteries, and shape the future of energy storage? This role sits where electrochemistry, data/ML, and control theory meet. You will develop the physics-informed and data-driven methods that estimate state of charge, health, power capability, and remaining life, and validate them against lab and fleet data until they are ready to ship. Your algorithms will decide how hard a battery can be pushed, when it needs attention, and how long it will last, so they shape the safety, warranty, and economics of the whole product. You will work side by side with cell, systems, and software engineers, and as an early hire you will help define how we develop and validate battery algorithms. This role can be based in Boston, Irvine, Sunnyvale, or Seattle, and involves occasional domestic and international travel, up to 10% of the time. Key job responsibilities • Develop and own algorithms for state of charge, state of health, power capability, and remaining useful life, combining physics-based models with data-driven and estimation methods. • Deliver validated reference implementations, test vectors, and accuracy targets that define what production code must achieve. • Design the test plans and data needed to develop and validate algorithms, working with the cell and testing teams to turn lab and field data into model improvements. • Analyze algorithm performance across the deployed fleet, find where models drift or fail, and drive improvements. • Own the accuracy and robustness of battery algorithms, and sign off that production implementations match the validated reference. Provide algorithm evidence for safety certification. A day in the life Your week moves between data, models, and experiments. One day you might be working out why a health estimate drifts on cells that have aged in the field, and the next deciding which model structure holds accuracy at cold temperatures. You could be designing a test with the cell team to expose a behavior your model misses, or reviewing an implementation with an applied scientist to confirm that what ships matches what you validated. You will regularly share results with software, systems, and cell leaders, and your findings will shape decisions well beyond the algorithm itself. About the team We are building Amazon's energy storage capability, providing power for advanced technologies while contributing to a stable grid. Our product is highly connected, yet secure and resilient by design. We own it end to end, from software, ML, and controls through system design to cell selection and supply, and on to a safety-certified product in the field. The organization operates like a startup inside Amazon: small, fast-moving, and accountable for our decisions. We work across hardware, software, and science, with suppliers and partners worldwide, and with Amazon infrastructure, procurement, legal, and finance. The organization is at an early stage and growing quickly. You will be among the first algorithm hires, so you will help shape how we design, build, certify, and ship, with a direct line of sight to the product and business decisions your work informs. This is a dynamic role that will regularly present new challenges as the product and organization evolve.
US, WA, Redmond
At Amazon, we’re inventing on behalf of customers, and with Amazon Leo, we’re redefining what global connectivity looks like. Our mission is to deliver fast, affordable broadband to unserved and underserved communities around the world through a constellation of low Earth orbit (LEO) satellites. Every system we build helps connect people to education, healthcare, opportunity, and each other. As a Data Scientist, you will be responsible for developing advanced analytics and machine learning solutions for user terminals. You will develop predictive models to proactively identify possible user terminal failures in the field. You will work in a collaborative environment with a multi-disciplinary team, including constellation, RF, antenna, silicon, algorithm, and software engineers. Key job responsibilities As a Data Scientist, you will develop analytic tools for a team developing current and future user terminals. Your responsibilities include: • Develop statistical and analytical tool to enable the regression decision from on-orbit and lab measurement of user terminals • Publish documents and create compelling visualizations and presentations to communicate insights to stakeholders • Create and manage datasets for continued pre-training and supervised fine-tuning of LLMs • Develop scalable visualizations for analysis of user terminal performance • Work closely with constellation, RF, antenna, silicon, algorithm, and software engineers to root-cause the failures using data as the primary tool • Drive consensus on metrics and analysis approaches to support product development strategy Export Control Requirement: Due to applicable export control laws and regulations, candidates must be a U.S. citizen or national, U.S. permanent resident (i.e., current Green Card holder), or lawfully admitted into the U.S. as a refugee or granted asylum. A day in the life As a Data Scientist in the LEO Customer Terminal Team, you will work daily with satellite constellation, algorithm, RF, antenna, silicon, hardware, and software teams in a collaborative environment. Your focus will be using data as an intelligence source to enable design decisions for the team. About the team The LEO Customer Terminal team is responsible for developing both outdoor and indoor devices that enable customers to access internet service via the LEO satellite network. We own the entire process from early prototypes through mass production, including requirements documentation, architecture definition, hardware development, algorithm development, and all integration and verification testing.
US, CA, San Francisco
The Amazon Center for Quantum Computing (CQC) is seeking an Applied Scientist to research and design cryogenic control hardware for large-scale quantum processors, with an emphasis on superconducting digital circuits. Controlling and reading out orders of magnitude more qubits than today's systems demands rethinking where classical control lives and what technology implements it; you will design and model superconducting digital logic (e.g., single-flux-quantum-family circuits) and related cryogenic electronics, evaluate them against the wiring, power, and thermal budgets of a scaled system, and co-design control approaches with the qubit devices they serve. You will carry designs from concept through simulation, layout, design review, and fabrication handoff, and your evaluations will inform which control-electronics directions the organization invests in. Key job responsibilities • Design and simulate superconducting digital circuits and cryogenic control/readout electronics for scalable qubit control • Develop and validate Josephson-junction circuit models, timing and margin analyses, and verification methodology for superconducting digital designs • Quantify system-level trade-offs — wiring count, power dissipation at each cryogenic stage, latency, footprint — across candidate control architectures, and communicate recommendations • Co-design control interfaces with qubit-device colleagues so control electronics and quantum devices are developed as one system • Deliver design artifacts through the organization's design-request and design-review processes, including tapeout-ready layouts and verification reports • Present results internally and, where appropriate, publish in peer-reviewed venues About the team The Enabling Technologies team develops technologies and relationships facilitating scaling, including alternative quantum processor architectures, superconducting qubit device packaging, and cryogenic control device electronics. We work alongside a co-located quantum processor design team focused on current-generation devices; the two teams share design and review processes, and moving ideas between them is an explicit part of our charter.
US, WA, Redmond
Amazon Leo is building a constellation of thousands of low Earth orbit satellites. It will bring fast, reliable internet to customers beyond the reach of existing networks. On the Guidance, Navigation and Control (GNC) team, you will design the algorithms, simulations and onboard autonomy that keep every satellite precisely pointed and stable. Your work lets the payload serve customers on the ground, and it keeps the fleet flying safely. This is satellite GNC at constellation scale. The constellation is launching and growing today, so you will see your work go from concept to orbit, and then fly on many satellites. In this role, you will: - Take your GNC algorithms from concept to orbit, and stay with them through integration, test and on-orbit operations. - Write the production C++ flight code for your own algorithms. The code that flies is the code you write. - Solve hard problems in precise pointing, payload cueing and fleet autonomy. - Work across design, simulation, test and flight operations, not a narrow slice of the lifecycle. **Export Control Requirement:** Due to applicable export control laws and regulations, candidates must be a U.S. citizen or national, U.S. permanent resident (i.e., current Green Card holder), or lawfully admitted into the U.S. as a refugee or granted asylum. Key job responsibilities - Design and analyze algorithms for state estimation, attitude determination and control, precise pointing and payload cueing against demanding performance requirements. - Implement these algorithms in production C++ flight software. Our GNC engineers own their algorithms all the way into flight, and partner with our embedded software engineers to integrate and test them. - Write concepts of operations for each phase of a satellite's life, from launch and orbit raising to service and deorbit, including failures and recovery. Balance what the payload needs against momentum, power, thermal and ground-contact limits. - Develop momentum management strategies with reaction wheels and magnetic torquers. Help select, size and place the GNC sensors and actuators for future satellites, and design the algorithms that use them. These include star trackers, magnetometers, sun sensors, IMUs and GNSS receivers. - Develop onboard autonomy, including Fault Detection, Isolation and Recovery (FDIR), so that the fleet operates safely at scale with minimal ground intervention. - Build and validate satellite and constellation models and simulations at several levels of fidelity, including 6-DOF simulation and Monte Carlo analysis. Verify closed-loop performance in hardware-in-the-loop (HITL) tests. - Support component environmental tests, functional and performance checkout, and subsystem and satellite integration in our labs. - Analyze telemetry from the whole fleet. Build the metrics, dashboards and tools that track pointing performance and payload operations time as the constellation grows, and use them to improve the fleet. - Operate the GNC system on orbit and take part in the team's on-call rotation. Review fleet data, support maneuvers, and resolve anomalies when on-orbit behavior differs from the ground model. - Review designs and code with the team, and document your algorithms, analyses and results clearly. A day in the life This role covers a wide range of problems across the spacecraft and the network. You will derive and prototype estimation and control algorithms, and then turn them into production flight code. You will prove them in 6-DOF and Monte Carlo simulation and in hardware-in-the-loop tests in the lab. You will support satellite integration. When a satellite does not behave as the ground model predicted, you will dig into fleet telemetry, find the cause and drive the fix. You will own each problem end to end, with the focus to go deep on it. GNC is at the center of safe flight, so you will work across the whole program. The payload, power, thermal and operations teams all depend on GNC, and you will balance their needs against the pointing and momentum budgets. You will partner closely with avionics, flight software and embedded software engineers to make your algorithms run reliably on the spacecraft. You will help the mechanical teams shape the next generation of satellite hardware, from reaction wheels to star trackers. As satellites move from design through manufacturing into orbit, you will work with systems engineering and Assembly, Integration and Test. The problems are hard: tight pointing requirements, a large constellation to keep flying safely, and the chance to invent capabilities that do not exist yet. Constant work with other teams makes this a place where you keep learning and growing. About the team You will join the Satellite Attitude Determination and Control team. We own GNC for our satellites end to end: algorithm design, flight software, lab integration, and flight operations of our own satellites on orbit. Our engineers bring deep experience from many satellite programs and other flight vehicles, with real strength in both our mission and the core GNC disciplines. We design, prototype, test, iterate and learn together. Because GNC is central to safe flight, we drive the concepts of operations and many of the system-level analyses on the program. We also have a lot of freedom. We trust our engineers to choose their own tools and approaches for the problems they face, and to own the results. If you want the autonomy to solve hard problems your way, with a team of strong engineers, you will feel at home here.
US, WA, Redmond
Amazon Leo is building a constellation of thousands of low Earth orbit satellites. It will bring fast, reliable internet to customers beyond the reach of existing networks. On the Guidance, Navigation and Control (GNC) team, you will design the algorithms, simulations and onboard autonomy that keep every satellite precisely pointed and stable. Your work lets the payload serve customers on the ground, and it keeps the fleet flying safely. This is satellite GNC at constellation scale. The constellation is launching and growing today, so you will see your work go from concept to orbit, and then fly on many satellites. In this role, you will: - Take your GNC algorithms from concept to orbit, and stay with them through integration, test and on-orbit operations. - Write the production C++ flight code for your own algorithms. The code that flies is the code you write. - Solve hard problems in precise pointing, payload cueing and fleet autonomy. - Work across design, simulation, test and flight operations, not a narrow slice of the lifecycle. **Export Control Requirement:** Due to applicable export control laws and regulations, candidates must be a U.S. citizen or national, U.S. permanent resident (i.e., current Green Card holder), or lawfully admitted into the U.S. as a refugee or granted asylum. Key job responsibilities - Design and analyze algorithms for state estimation, attitude determination and control, precise pointing and payload cueing against demanding performance requirements. - Implement these algorithms in production C++ flight software. Our GNC engineers own their algorithms all the way into flight, and partner with our embedded software engineers to integrate and test them. - Write concepts of operations for each phase of a satellite's life, from launch and orbit raising to service and deorbit, including failures and recovery. Balance what the payload needs against momentum, power, thermal and ground-contact limits. - Develop momentum management strategies with reaction wheels and magnetic torquers. Help select, size and place the GNC sensors and actuators for future satellites, and design the algorithms that use them. These include star trackers, magnetometers, sun sensors, IMUs and GNSS receivers. - Develop onboard autonomy, including Fault Detection, Isolation and Recovery (FDIR), so that the fleet operates safely at scale with minimal ground intervention. - Build and validate satellite and constellation models and simulations at several levels of fidelity, including 6-DOF simulation and Monte Carlo analysis. Verify closed-loop performance in hardware-in-the-loop (HITL) tests. - Support component environmental tests, functional and performance checkout, and subsystem and satellite integration in our labs. - Analyze telemetry from the whole fleet. Build the metrics, dashboards and tools that track pointing performance and payload operations time as the constellation grows, and use them to improve the fleet. - Operate the GNC system on orbit and take part in the team's on-call rotation. Review fleet data, support maneuvers, and resolve anomalies when on-orbit behavior differs from the ground model. - Review designs and code with the team, and document your algorithms, analyses and results clearly. A day in the life This role covers a wide range of problems across the spacecraft and the network. You will derive and prototype estimation and control algorithms, and then turn them into production flight code. You will prove them in 6-DOF and Monte Carlo simulation and in hardware-in-the-loop tests in the lab. You will support satellite integration. When a satellite does not behave as the ground model predicted, you will dig into fleet telemetry, find the cause and drive the fix. You will own each problem end to end, with the focus to go deep on it. GNC is at the center of safe flight, so you will work across the whole program. The payload, power, thermal and operations teams all depend on GNC, and you will balance their needs against the pointing and momentum budgets. You will partner closely with avionics, flight software and embedded software engineers to make your algorithms run reliably on the spacecraft. You will help the mechanical teams shape the next generation of satellite hardware, from reaction wheels to star trackers. As satellites move from design through manufacturing into orbit, you will work with systems engineering and Assembly, Integration and Test. The problems are hard: tight pointing requirements, a large constellation to keep flying safely, and the chance to invent capabilities that do not exist yet. Constant work with other teams makes this a place where you keep learning and growing. About the team You will join the Satellite Attitude Determination and Control team. We own GNC for our satellites end to end: algorithm design, flight software, lab integration, and flight operations of our own satellites on orbit. Our engineers bring deep experience from many satellite programs and other flight vehicles, with real strength in both our mission and the core GNC disciplines. We design, prototype, test, iterate and learn together. Because GNC is central to safe flight, we drive the concepts of operations and many of the system-level analyses on the program. We also have a lot of freedom. We trust our engineers to choose their own tools and approaches for the problems they face, and to own the results. If you want the autonomy to solve hard problems your way, with a team of strong engineers, you will feel at home here.
US, WA, Seattle
Do you want to make AI provably correct for chip design and verification? We're hiring an Applied Scientist to combine large language models with formal methods such as Lean 4 and turn AI from a helpful assistant into a trusted engineering tool for silicon. Our team builds the formal verification methods behind custom silicon at Amazon scale. As AI agents take on more of chip design and verification, we need people who can make sure their output is checkable, sound and grounded in mathematics. You will research and build systems where LLMs work alongside interactive theorem provers, SMT solvers and model checkers. Engineers write the specifications. AI refines them against RTL, generates proof steps and closes proofs at machine speed. Every result is checked by a formal kernel, not trusted on faith. Work at the frontier of AI reasoning, theorem proving and hardware verification Ship solutions that directly affect real chips running in AWS data centers Publish and help shape how the industry formalizes AI use in hardware engineering Key job responsibilities Design and build LLM-driven systems for formal verification and design tasks, including proof generation, invariant and lemma discovery, spec-to-RTL refinement, and counterexample analysis Develop frameworks in Lean 4 (and related tools such as SMT solvers, model checkers and commercial FV engines) that formally check AI-generated artifacts, so they are correct by construction rather than just plausible Formalize hardware semantics, ISAs and microarchitectural properties in Lean 4 to create machine-checkable foundations for AI agents to reason over Track and evaluate state-of-the-art AI work (LLM agents, autoformalization, neural theorem proving, reinforcement learning from proof feedback) and turn promising research into production tools Build benchmarks and evaluation methods that measure AI soundness, proof success rate and engineering impact on real designs Partner with FV and design engineers to find the highest-leverage problems and deploy solutions into active chip programs Publish research, file patents and represent Amazon in the formal methods and AI communities Mentor engineers and scientists on applying formal AI methods
US, CA, Pasadena
We are seeking Applied Scientists to join Amazon. In this role, you will develop and deploy the science of safe autonomy on highly dynamic robots, advancing the frontier through both foundational research and realization on hardware. You will contribute to achieving safe autonomy across three key areas: control barrier function (CBF) theory for robust and performant safety on hardware; safe reinforcement learning for agile whole-body control; and layered safety filters that interface with learning, perception, and semantic reasoning systems. You will validate your work experimentally on state-of-the-art robotic platforms — removing bottlenecks to deployment and enabling robots to safely operate around humans. You will work with the inventor of CBFs, as well as top scientists and engineers at Amazon developing the next generation of safe autonomy. Key job responsibilities -Advance the science of safe autonomy from formal foundations to the integration with learning and perception to validation on hardware — with particular emphasis on methods that bridge these domains. -Test and validate hardware, with a focus on next generation robotic systems. Including locomotion, manipulation and loco-manipulation. -Leverage the deployments on hardware to validate the underlying science, identifying gaps between theory and practice that drive the next cycle of research. hardware to validate the underlying science, identifying gaps between theory and practice that drive the next cycle of research. -Publish research at top-tier robotics, control, and ML venues, and contribute to Amazon’s scientific reputation in advanced robotics. -Collaborate with SAF Lab and Amazon production teams to move research into robots deployed at Amazon scale. A day in the life Amazon offers a full range of benefits that support you and eligible family members, including domestic partners and their children. Benefits can vary by location, the number of regularly scheduled hours you work, length of employment, and job status such as seasonal or temporary employment. The benefits that generally apply to regular, full-time employees include: -Medical, Dental, and Vision Coverage -Maternity and Parental Leave Options -Paid Time Off (PTO) -401(k) Plan If you are not sure that every qualification on the list above describes you exactly, we’d still love to hear from you! At Amazon, we value people with unique backgrounds, experiences, and skillsets. If you’re passionate about this role and want to make an impact on a global scale, please apply! About the team The Safe Autonomy Frontiers (SAF) Lab is the first industry research lab dedicated to safe autonomy, founded by the inventor of control barrier functions. We are developing a universal safety layer for the next generation of robotic systems: mobile robots, manipulators, quadrupeds, and humanoids. Here you will advance performant safety for highly dynamic robots — CBF theory integrated with perception and learning, evaluated on next-generation platforms — and your work will underpin robots operating alongside people at Amazon’s unprecedented scale.
US, MA, North Reading
Robotics Applied Scientist II Intern/Co-op (2027) Are you passionate about robotics and research? Do you want to solve real customer problems through innovative technology? Do you enjoy working on scalable research and projects in a collaborative team environment? Do you want to see your science solutions directly impact millions of customers worldwide? We’re looking for current PhD students with a passion for robotic research and applications to join us in 2027 as Robotics Applied Scientist II Interns/Co-ops to shape the future of robotics and transform automation at Amazon’s scale. By applying to this role, you will be considered for Robotics Applied Scientist II Intern/Co-op (2027) opportunities across various Robotics teams at Amazon with different robotics research focus. Internship positions are available for multiple locations, durations (3 to 6+ months), and year-round start dates (winter, spring, summer, fall). Join us at the forefront of Applied Science, where your contributions will shape the future of Robotics! Key job responsibilities As a Robotics Applied Scientist II Intern/Co-op, you will work alongside global experts to advance the state of the art in robotics and artificial intelligence, tackling innovative problems that bridge the gap between research and real-world robotics implementation to innovate and build solutions on behalf of our customers. We are particularly interested in students with specialization and expertise in one or more of the following research areas: - Robotics - Artificial Intelligence - Robotics Manipulation (e.g., robot arm, grasping, dexterous manipulation, end of arm tools/end effector) - Autonomous Mobile Robots - Mobile Manipulation - Multi-Robot Coordination - Autonomous Navigation - Locomotion - Motion/Path Planning - Controls - Perception - Sensing - Robot Learning - Machine Learning - Computer Vision - Large Language Models (LLM), Vision-Language-Action Models (VLA), Vision-Language-Models (VLM) - World Models - Human-Robot Interaction - Robotics Simulation - Optimization - And related! Key responsibilities may include: - Develop novel, scalable algorithms and/or modeling techniques that advance the state-of-the-art in science areas such as robotics, artificial intelligence, or related - Tackle challenging, innovative research problems within robotics, artificial intelligence, or related that bridge the gap between theoretical advancement and practical robotics implementation - Collaborate with cross-functional teams to solve complex problems, and to communicate research findings clearly, leveraging your expertise in robotics, artificial intelligence, or related - Demonstrate the ability to work independently, thrive in a fast-paced, ever-changing environment - Leverage AI-powered tools where applicable to accelerate research, experimentation, and prototyping. - Critically review and validate outputs from AI tools and automated systems. About the team Explore Robotics Teams at Amazon At Amazon, robotics is where innovative research meets real-world deployment. There are many different specialized robotics teams, including the few highlighted below. Learn more about Robotics at Amazon here: https://www.amazon.science/research-areas/robotics Fulfillment Technologies and Robotics We build dynamic partnerships between people and intelligent machines. This intricate collaboration helps Amazon fulfill orders with unmatched accuracy. Working in symphony with our robotic technology, you’ll have the opportunity to extend your technical capabilities by working alongside some of the industry’s most advanced technologies. This includes our fleet of autonomous mobile robots, sophisticated control software, and technologies like language perception, machine learning, object recognition, and semantic understanding of commands. These technologies help us deliver an ever-improving customer and employee experience, as well as improve the safety of our facilities. Amazon Robotics Amazon Robotics is at the epicenter of robotics innovation. We design, test, and build mobile robots and mechatronic systems. Our offices, labs, test floors, and factories are close together, which means we can design the hardware, write the code, test, build, and watch it all come to life. Our areas of focus include advancements in autonomous mobility, AI, robotic management software, and much more. Fauna Fauna is focused on building capable, safe, and genuinely delightful robots for everyday life. Fauna’s goal is simple: make robots people actually want to live and interact with everyday. Fauna develops tightly integrated hardware and software systems that make it faster, safer, and more intuitive to create real-world robotic products. Fauna’s work spans the full stack: mechanical design, control systems, dynamic modeling, and intelligent software. The focus is not just functionality, but overall experience. Amazon Fulfillment Technologies Amazon Fulfillment Technologies develops tech solutions that create network optimizations for all Amazon sites worldwide. We own network infrastructure and devices in all buildings and labor planning and scheduling technology for fulfillment centers in North America and Europe.
US, WA, Seattle
Amazon Web Services (AWS) provides a highly reliable, scalable, low-cost infrastructure platform in the cloud that powers hundreds of thousands of businesses in 190 countries around the world. We are hiring an Applied Scientist in our AWS Sales Science & Economics team. The team uses machine learning, econometrics, and data science to optimize AWS’s service catalogue, driving customer engagement and generating insights to guide AWS sales strategy. We use detailed customer behavioral and usage data to predict and understand what customers want from AWS. Key job responsibilities As an Applied Scientist, you will apply the latest research, state of the art machine learning and GenAI to build highly scalable models in support of predictive research and products used across AWS Sales and leadership. You will be responsible for designing and executing science experiments in short and agile cycles. You will use the latest AWS technologies in Machine Learning, Data Science, and AI. You will work closely with business leaders to identify opportunities to drive business growth by applying machine learning and building out predictive models to support analytics and insights. The ideal candidate will be curious and able to learn new technologies quickly. You will have hands-on experience building data science and machine learning pipelines in production working with large scale data sets. You are passionate about sharing your knowledge to enable others. You empathize with the needs of our customers and align your work accordingly, and provide thought leadership and guidance across the organization. A day in the life You will start your day reviewing experimental results and refining model architectures based on new insights. You might spend the morning prototyping a novel approach to a customer-facing problem, then shift to a design review with engineering partners in the afternoon to align on how your solution integrates into production systems. Along the way, you will mentor junior scientists, contribute to peer reviews, and stay connected to the latest research through reading groups and internal science discussions. Your contributions will move from research notebooks to live systems that customers interact with every day. About the team Our team is focused on applying advanced science and machine learning to build intelligent systems that improve the Amazon experience. We value collaboration, intellectual rigor, and a willingness to explore new ideas. You will work alongside scientists and engineers who are committed to turning ambitious research into practical, high-impact solutions. We are investing in the next generation of machine learning capabilities and are looking for team members who are excited to help define that future.