What this covers

This is a three-part talk from Bridgewater Associates' applied AI team about PAT, an internal analyst tool used by hundreds of the firm's investors. Brendan McManus frames the project, Michael Ran demos it live on a question about oil supply shocks and the Middle East conflict, and Santi Weight explains the architecture. McManus starts from the firm's history: a 1980 bond system written on a yellow legal pad, with traders recording the exact causal logic of each decision, and fifty years of encoding markets into compounding systems. He places PAT inside the research circle, and says it targets the investigative stage that takes human analysts days or weeks, not trading itself. He also argues that the firm had to build it through a startup-like internal team of investors, technologists and scientists working side by side.

The demo covers several concrete design choices. PAT's time series search improved from roughly 50% to 90% accuracy once it inspected candidates the way a human does, checking frequency, currency and plausibility against priors, not just names. Each investor gets a separate PAT with its own context and tools, so access matches what that person may see. PAT writes its outputs back into the same database it reads from, so agent and human work can build on each other. It asks clarifying questions up front because "the plan is the analysis." A Teach button then generates benchmarks and pull requests to fix its own mistakes.

Weight's section treats code generation as a compiler problem. A detailed plan lets code be generated in parallel, so a 3-dataframe analysis and a 30-dataframe one take about the same time. Validation agents run in a graph, and the orchestration is ordinary Python, so validation is forced rather than left to an agent's discretion. Static analysis and caching make small edits nearly instant. He compares this with Claude Code, which is faster on a first run and has a larger advantage on re-runs. The talk closes with three takeaways: specialize agents, benchmark narrow workflows, and think like a compiler.

Sharpest takeaway

Bridgewater built PAT (Pocket Analyst Tool), an AI analyst that automates expert research work by decomposing human investor workflows into specialized agents with deterministic code generation, enabling hours of analysis in minutes while learning from every interaction.

  • PAT decomposes the research circle into discrete sub-agents, each handling one phase (perception, questioning, investigation, synthesis) of what human investors do
  • Deterministic code generation through detailed planning and compiler-like validation enforces correctness rather than relying on LLM vibes
  • Continuous learning loop from hundreds of internal users generates benchmarks that improve PAT for all users, creating a compounding flywheel

The claims · ranked34 claims · weighted by value

This asset isn't compiled yet

You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.

0.69

The last step of the research circle—feeding learned insights back into shared expert systems—is key because everything investors learn becomes part of a shared memory that other humans can draw upon.

causalhigh valueestablishednovelty 1/4durability 3/4· Brendan McManus

“It's really this last step that's key. Everything that we learn as a function of this process gets put back into a shared memory that humans are able to draw upon.”

0.68

Bridgewater has spent 50 years encoding markets and economies into compounding systems, starting with a 1980 bond system written on a yellow legal pad where traders would write down exact causal logic for every trade decision.

factualhigh valueestablishednovelty 0/4durability 4/4· Brendan McManus

“Bridgewater has been spending decades — 50 years — thinking about how to encode markets and economies into compounding systems. And all of this really started with what you're seeing on the screen: our 1980 bond system, written down on a yellow legal pad. The idea here was quite simple. Every time you wanted to make a trade, write down exactly the rules for why you think that trade makes sense.”

0.52

Multi-archetype teams with investors, technologists, and scientists sitting side by side and building together are necessary for building AI systems for expert users, because investors bring context and domain expertise, technologists bring architectural capability, and scientists bring rigor.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Brendan McManus

“In addition to this, we've established multi-archetype teams where we have investors, technologists, and scientists sitting side by side, building alongside each other. And I think this is really a key thing if you're going to build out a product like this. Investors bring the context and the domain-specific expertise. Technologists bring the architectural capability. And scientists bring the rigor.”

0.52

PAT outputs from analyses are placed into the same time series database that inputs are pulled from, which means PAT-generated outputs are indistinguishable from human-generated series and can serve as inputs to subsequent analyses, enabling humans and agents to compound each other's work.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Michael Ran

“And before we proceed from execution, I want to call out the fact that the time series outputs from these analyses land in the same database that we actually pulled our inputs from. And I think this is important for a couple of reasons. One: it shows that any output from a PAT analysis is indistinguishable from any of the human-uploaded series that we've been producing for many years. And more importantly, any output from a PAT analysis can serve as an input to a subsequent one. So this creates an environment where humans and agents can very easily compound and leverage each other's work.”

0.52

Bridgewater's key takeaway is to specialize agents rather than build generic, powerful agents—while specialized agents make less impressive demos, they are much harder to make into daily workflows that users can depend on.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Santi Weight

“The first takeaway: we really believe in specializing our agents. We don't really believe in generic, powerful agents. They make really cool demos. I'm sure many of us have given them. But it's really hard to make that a daily workflow that you can depend on.”

0.52

Bridgewater allows investors to contribute to PAT's context like a developer on a codebase, because users are often better at writing context than the builders, and maintaining low ego about contributions improves product quality.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Santi Weight

“We also let our investors contribute — Michael was showing that — contribute like a developer on the codebase. And this is a pretty good point. Your user is better at writing context than you are, most likely. And so having low ego about it and letting them contribute directly is a good way to win.”

0.52

Building AI tools like PAT at an established hedge fund requires being willing to shake yourself up by creating an internal applied AI startup incubated within the broader firm with the ability to move flexibly and quickly while drawing on broader firm resources.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Brendan McManus

“Well, it really starts with being willing to shake yourself up. The team that built out PAT was ultimately an internal applied AI startup that was incubated within the broader firm, with the ability to move incredibly flexibly and quickly while being able to draw upon the resources of the broader firm.”

0.52

Unlike systems like Claude Code where everyone interfaces with the same system prompt and tools, each person at Bridgewater has a unique version of PAT tailored to what they can and cannot see, which is achieved through unique context and tool sets for each person's PAT.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Michael Ran

“So unlike a harness like Claude Code, where everyone is interfacing with the same system prompt and the same tools, each person at Bridgewater has a unique version of PAT tailored to what they can and cannot see. Practically speaking, this uniqueness is just a function of what context and what tools each person's PAT has.”

0.52

When a human researcher searches for data, they don't just anchor to the name of the time series—they also look at frequency, currency, and whether values align with their priors; embedding this human-like inspection logic into PAT's search agent increased accuracy from roughly 50% to 90%.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Michael Ran

“When a human researcher is looking for data, they don't just anchor to what the name of the time series is. They'll look at the frequency, the currency of the series, and then most importantly, whether the values in the series align with their priors. So embedding this sort of reasoning into our search agent is something that got us up from roughly 50% accuracy all the way to 90%.”

0.52

During PAT's development, the team came to view that the plan really is the analysis—if they can create a high-quality, detailed plan, they are confident they can consistently execute it and produce desired outputs.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Michael Ran

“During PAT's development, we came to the view that the plan really is the analysis. If we can create a high-quality, detailed plan, we felt confident that we could consistently take this plan, intelligently execute it, and produce our desired outputs.”

0.52

Separating the chat agent from the coding agent allows Bridgewater to keep chat focused purely on investment content (not programming), because investors are not programmers by trade—the result is that coding becomes a pure implementation detail invisible to users.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Santi Weight

“Why would you separate your two agents? Early on, we decided that our investors are not programmers by trade — they care about investment. So we decided to keep the chat purely about investment content. And the result is that we have a product where coding is a pure implementation detail. From the chat, you can't tell that there's code under the hood.”

0.50

Santi's background in compiler theory and programming language design is relevant because compilers and coding agents have similar requirements: they must be fully deterministic, fully correct, and reliable—you cannot have off-by-one errors in critical systems like aviation.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Santi Weight

“My background is in compiler theory and programming language design. Compilers have a very similar footprint of requirements: they're fully deterministic, fully correct, and reliable. You can't have an off-by-one error when you're flying a plane, for example. And there's a similar shape here. A compiler takes user code and compiles it down to something like JavaScript, and a coding agent takes a user prompt or plan and compiles it down to Python.”

0.50

Because PAT is a reproducible agent, Bridgewater can use something much more dependable than vibes-based or LLM-as-judge evaluations; instead they use deterministic benchmarking and testing.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Santi Weight

“And then because we have such a reproducible agent, as we're scaling and hill climbing and evaluating, we have something much more dependable than vibes-based or LLM-as-judge evals.”

0.48

PAT's interactive report uses the same in-house charting library and existing text deck that Bridgewater has developed internally for decades, enabling users to zoom, pan, and send data to the internal charting tool for further modifications.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Michael Ran

“And then the result itself is this interactive report where the visualizations look just like the ones investors at Bridgewater are producing. PAT is using the same in-house charting library and leveraging the same existing text deck that we've been developing internally for decades. So you can see here how you can zoom in, zoom out of these charts, and you can even send the data from this interactive report to our internal charting tool, where you can make further modifications on the fly.”

0.48

PAT is focused on the investigative portion of the research process—the kinds of deep analytical work that take human analysts days and weeks to perform—rather than on how Bridgewater trades.

definitionhigh valuespeaker onlynovelty 1/4durability 3/4· Brendan McManus

“We called this tool PAT, the Pocket Analyst Tool. And just to set expectations: PAT is not about how we trade. It's really about performing deep exploratory research, enabling our investors to go after questions that they never would have had the bandwidth to pursue before.”

0.48

PAT must use all the proprietary tools that human analysts have access to—including tools for visualization, diagnostics, and evaluating the quality of indicator ideas—to truly replicate human analytical workflows.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Brendan McManus

“PAT also has to be able to use all of the different tools that our human analysts have access to. All of the proprietary tools that we've built out for visualization, for diagnostics, or for evaluating the quality of the indicator ideas that we've built out.”

0.48

PAT must learn and compound upon its own learning not just for one investor, but for every investor at Bridgewater, creating a continuous improvement mechanism.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Brendan McManus

“And lastly, PAT has to learn. It has to compound upon its own learning, not just for one investor, but for every investor at the firm.”

0.48

Having hundreds of expert users internally using AI tools daily generates continuous signal on how those tools should evolve and improve over time.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Brendan McManus

“And speaking of expert users, we have hundreds of them internally using all of our AI tools — not just PAT — generating signal daily on how these things should be evolving and improving over time.”

0.48

Bridgewater's investment domain context for PAT is high quality because they teach the chat agent how to talk like a Bridgewater investor, including extensive jargon, enabling investors and the agent to communicate like coworkers.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Santi Weight

“What we're going to talk about now is that our investment domain context is really high quality. We teach the chat agent how to talk like a Bridgewater investor. There's a lot of jargon to teach it. And so the user and the agent both talk to each other like coworkers at Bridgewater.”

0.48

The instantaneous code execution on iterative modifications means investors can make small tweaks to their analysis without overhead, enabling faster exploration and hypothesis testing.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Santi Weight

“And this means that when you're working on the product as an investor, you can make small tweaks to your investment analysis without the overhead of a regular iteration.”

0.47

PAT asks clarifying questions to help humans—who tend to under-invest in planning—flesh out a high-quality research plan before beginning analysis.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Michael Ran

“So while agents asking clarifying questions is pretty standard in chatbots nowadays, the thing we really focused on was the substance of those questions. We devoted lots of time and energy into developing the context and benchmarks that helped shape this capability. We taught PAT what makes a good research question versus a bad one, so this back-and-forth helps humans — who tend to under-invest in planning — flesh out what we consider to be a high-quality research plan.”

0.47

Benchmarking PAT against Claude Code shows that on the first run, PAT is faster due to never double-loading data or double-executing intermediates, but the real win appears on the second run when the last chart in the plan is modified with just a name change.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Santi Weight

“This is an example benchmark we have for Claude Code invoking its own code versus Pocket Analyst. We're faster because we never double-load data or double-execute intermediates. But the actual win is not on the first time you run the code, it's the second time. Oh, there we go. This is a benchmark where we take the last chart in a plan and just change the name. Claude Code is going to go and rerun all the code — basically the same amount of time, though it is faster at editing the code. But Pocket Analyst basically has instantaneous code execution for the second round.”

0.47

With a detailed plan specifying data frame dependencies and schemas, code generation can happen in parallel across sub-agents, such that a 3-data-frame analysis and a 30-data-frame analysis take roughly the same amount of time to generate code.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Michael Ran

“So now we have the plan locked in, and the first step of plan execution is generating the code. Given we have such a detailed plan, we're able to generate the code for each data frame in our analysis in parallel using sub-agents. And this works because each sub-agent knows what data frames it depends on, knows the schemas of those data frames, and also knows what the schema of the data frame it's producing should be. So this lets an analysis with three data frames or a medium-sized one with something like 30 data frames take roughly the same amount of time to generate code in both scenarios.”

0.47

Because of detailed task specifications, the analysis plan is not just a to-do list like Claude Code, but rather a natural language Python project, enabling application of advanced code generation techniques

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Santi Wei

“The result is that our analysis plan isn't just a to-do list like you'd see in Claude Code. Instead, we think of it as a natural language Python project. And because we have so much detail, now we're going to go into code gen and we can apply some fancy techniques.”

0.47

PAT enforces correctness through architecture: no agentic orchestration, only regular Python code, so guardrails are hard and agents cannot forget to validate—they are forced to validate.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Santi Weight

“The main point I want you to take away from this part of the slides is that we enforce correctness in the architecture. Again, no agentic orchestration. This is regular Python code, so the guardrails are really hard, and the agents cannot forget to validate. They are forced to validate.”

0.45

There are two primary ways PAT gets better: autonomously through agents reviewing completed conversations for improvement opportunities, and explicitly through users clicking a Teach button if they think PAT should have suggested something from the start.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Michael Ran

“There are two primary ways in which this happens. The first is the autonomous one that Brendan alluded to earlier, where we have agents reviewing completed conversations looking for ways in which PAT can get smarter. And the second one — and the one we'll be showing here — is the more explicit option, where a user can kick off this learning process within the context of an analysis if they think there is something to be learned from their interaction with PAT.”

0.45

After execution, PAT double-checks its work like a junior analyst would—reviewing computed data and visualizations for sensibility, diagnosing and refining if something looks off, and only returning results once satisfied.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Michael Ran

“Now, after execution completes, just like you'd want your junior analyst to double-check their work before coming back to you, we want PAT to do the same. So at this point in the analysis, PAT will look at the data that's been computed and the visualizations it's produced and make sure that the numbers seem sensible and the charts look clean. If it sees something that looks a bit off, PAT will take a step back, diagnose the issue, refine the analysis, and make sure it's satisfied with its result before coming back to the user.”

0.45

As investors, Bridgewater must deeply understand every major dynamic shaping global markets and economies—including AI itself—by analyzing the shape of AI supply-demand mismatches, the buildout drivers, and their ultimate market impact, which is table stakes for their investment process.

normativehigh valuespeaker onlynovelty 1/4durability 2/4· Brendan McManus

“As investors, we need to deeply understand every major dynamic that's shaping global markets and economies. Just as we had to understand COVID or the recent oil supply shock, we've also had to understand AI. What's the shape of the supply-demand mismatch? What's driving the buildout? And how do these things ultimately impact markets? This understanding is table stakes for us as investors.”

0.43

The chat agent has access to tool calls for data series search, unstructured search, and once it has gathered the necessary data, it invokes a sub-agent (the coding agent) via a tool call to produce a Python-Pandas analysis.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Santi Weight

“The chat agent has access to tool calls. Michael called some of these out. So there's data series search, unstructured search, and each of these would be a tool call themselves. And then once the chat agent knows all the data it needs for its analysis, it's going to make a plan and invoke a tool call to a sub-agent, which is our coding agent. And that's going to produce a Python-Pandas analysis.”

0.43

Coding agents are difficult to build because they are fickle, unpredictable, often make mistakes, and in the worst cases can go off the rails and attempt to damage data; building them into daily workflows at a hedge fund trading billions of dollars requires systems that cannot rely on vibe-coded analysis.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Santi Weight

“I'm sure many of you in the crowd are building coding agent products just like we are. And it's really hard to build one. Coding agents are really fickle and unpredictable, they often make mistakes. And when you're really unlucky, they'll go off the rails and try to nuke your data and all the rest of it. And so it's already hard to build a good product that people like, but then it's really hard to build a product they want to embed into their daily workflows. At a hedge fund, we're trying to trade billions of dollars, and so we can't have just vibe-coded analysis be the underpinning of how these analyses go.”

0.43

When a user submits a teaching interaction, a front-end agent creates a benchmark that PAT should fail, confirming that the poor behavior can be reproduced, then iterates on context/harness until the benchmark passes, and finally confirms no regression across the broader test suite before submitting a pull request.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Michael Ran

“an agent on the back end will create a benchmark that we expect to fail. This shows that we can reproduce this poor behavior. It will then iterate on our context repositories or the harness itself until that benchmark passes. And then after confirming that by making this benchmark pass, we didn't cause the rest of our suite to fail, we get a Slack message with the pull request, including the changes the agent wants to make to PAT.”

0.43

PAT's code execution strategy is: take the task from the plan and the code generated for it, run the code, compare output to the task specification, and if incorrect, edit the code and iterate until correct—LLMs are not one-shot perfect on these tasks.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Santi Weight

“Instead, what we do is take the task that came in from the plan and the code that was generated from the task, then run the code, compare it to the task, see if it's correct, and if it's not, edit the code and keep going until we're done.”

0.43

PAT's first step in analyzing a question is searching web and unstructured internal data (millions of documents including broker pieces, earnings transcripts, internal emails, updating in near real time with thousands of new pieces daily) to understand what's happening in the world contextualized against history.

factualhigh valuespeaker onlynovelty 1/4durability 3/4· Michael Ran

“The first thing that takes place in this analysis is PAT searches the web and our vault of unstructured data to better understand what's happening in the world today, contextualized against history... We have a database containing millions of documents from all around the world that includes content such as broker pieces, earnings transcripts, internal emails, and more. And this database is updating in near real time, with thousands of new pieces coming in each day.”

0.17

The chat agent in PAT is implemented in LangGraph, primarily for persistent state management, and it has out-of-the-box support for cancellations and continuations, which is better than the custom state management Bridgewater previously built.

factualspeaker onlynovelty 0/4durability 2/4· Santi Weight

“The chat agent is implemented in LangGraph. We primarily use it for persistent state. It has out-of-the-box cancellations and continuations. We used to manage that stuff ourselves with a much worse effect.”