Author: Lincoln Wang | Founder of MindsLeap | Global Partner at Founders Space | Founder of Founders AI Club
A Yellow Piece of Paper from 1980
Brendan McManus from Bridgewater didn't start with models or frameworks in this LangChain talk.
He started with a piece of paper.
It was Bridgewater's 1980 bond system, written on yellow legal paper. By his account, the company has spent the past fifty years doing something very笨 and very difficult: writing down their understanding of markets and the economy, rule by rule.
He quoted a line from Ray Dalio:
"Every time you want to make a trade, write down the rules you believe make that trade work."
On the surface, this is about investment methodology. But viewed today, it's really about how organizations prepare for the AI era.
Because while most companies are still asking "can we plug in a model?", Bridgewater is already standing in a very different position. It's not hastily补课 for AI agents — it has turned fifty years of accumulated research methods, data systems, tools, terminology, and decision logic into context that AI agents can call upon.
Brendan was direct:
"We don't need to go back and rewrite everything for AI agents — it's already there."
This may be the most important signal from this talk.
AI doesn't make organizations smart out of nowhere. It amplifies what an organization has already written down, structured, and made reusable.
PAT Is Not a Chatbot
The tool Bridgewater built is called PAT — Pocket Analyst Tool — essentially an internal AI analyst.
Brendan set boundaries from the start:
"PAT is not about how we trade."
That's important. Because hearing Bridgewater, AI, and investing together, many people immediately think of automated trading, strategy prediction, or some mysterious financial black box. But what they showcased this time isn't about trading decisions — it's about the research process.
PAT solves a very specific problem: investors have a huge number of questions worth asking every day, but human time is limited, and many deep exploratory research tasks don't have enough bandwidth. PAT's goal is to compress research that used to take experts hours into something that can be advanced in minutes.
In the demo, they had PAT analyze the market reaction to a recent Middle East conflict and compare it with historically similar oil supply shocks. This sounds like a normal research question, but once it lands inside a real institution, it becomes much more complex.
It needs to find the right time series data, understand which historical events are comparable, generate charts, check whether numbers are reasonable, and leave the analysis process transparent so both humans and backend AI agents can diagnose it.
This isn't "asking a model a question."
This is having an AI agent enter a real organization's research workflow.
The Real Moat Is Context
PAT works not because it connects to more models than others, but because Bridgewater has context that others can hardly replicate.
Brendan said PAT must read structured data — like decades of stock prices and macro time series — and also unstructured materials like external research reports and internal memos. It also needs to use proprietary tools that human analysts already use, including visualization, diagnostics, and metric evaluation tools.
More critically, PAT knows how analysts should work.
"It knows exactly how our analysts should conduct research, because we've been writing that down for fifty years."
This is especially worth noting for Chinese entrepreneurs.
Many enterprises today, when advancing AI agents, their first instinct is to find models, find tools, find vendors. But what truly determines the ceiling of an agent often isn't model parameters — it's whether the enterprise has clearly documented its own business methods.
Has a sales team documented how to judge customer quality?
Has a manufacturing company documented how to investigate anomalous orders?
Has a consulting firm documented its methods for interviews, diagnostics, delivery, and retrospectives?
If these things exist only in the heads of a few veteran employees, AI agents can only become smart external consultants who don't understand your company. They might answer general questions, but they'll struggle to own real processes.
Bridgewater's case reminds us: the first asset of AI-native organizations isn't prompts — it's organizational knowledge that can be read by machines, reviewed by humans, and continuously updated.
"The Plan Itself Is the Analysis"
One of my favorite lines from this demo came from Michael Ran.
He said:
"The plan itself is the analysis."
This sounds like an engineering detail, but it's actually a deep working methodology judgment.
Before executing research, PAT doesn't rush to write code, pull data, or make charts. It first asks clarifying questions, turning a vague task into a high-quality research plan. Michael said humans often underestimate the importance of planning, so PAT's back-and-forth questioning actually helps people articulate the real problem.
This is exactly the opposite of how many enterprises use AI today.
Many people treat AI as an instant answer machine — throw in a question and immediately get a conclusion. But in complex business, the most valuable part often isn't the answer — it's getting the question right.
PAT's planning phase defines which data tables to generate, what structure each table should have, and how they connect. This planning takes time, but Bridgewater is willing to pay this cost because the execution afterward can proceed in parallel.
With a sufficiently detailed plan, AI agents can delegate code generation for different data tables to multiple sub-agents for parallel processing. Santi Weight later mentioned that in their architecture, a twenty-task plan can take nearly the same time as a three-task plan.
The insight is clear: AI agents aren't a single "smart brain" — they're an execution system that can decompose work, parallelize, verify, and reuse.
The future of enterprise process redesign likely isn't about giving one job to one AI — it's about breaking a complex process into many plannable, verifiable, parallelizable work units.
From 50% to 90%, Not by Magic
There's a very specific detail inside PAT: time series search.
Bridgewater has tens of millions of time series data internally. When a user asks a research question, PAT needs to find the right data. At first, with just basic search, accuracy was about 50%. Then they embedded the way human researchers check things — like data frequency, currency, prior knowledge, and data meaning — and accuracy improved to about 90%.
This detail is extremely valuable.
It shows that AI agent improvement doesn't come from saying "make the model smarter" — it comes from making the implicit actions of experts explicit.
A great analyst doesn't just look at keywords. They judge whether the data is from the same caliber, the same time frequency, and truly relevant to the question. Bridgewater turned these judgments into part of the search agent, and the system capability jumped a level.
This is the most easily overlooked part of enterprise AI adoption.
We always want to find a model with stronger general capability, but in real business, enormous value comes from details: which fields an approval should check, which signals a customer service escalation should judge, which seasonal factors to exclude for a store anomaly.
For AI agents to get stronger, these details must be written into the system.
A System That Can Be Taught
PAT isn't a one-and-done product.
Brendan said:
"PAT is no longer a prototype today."
It's been deployed internally at Bridgewater for months, with hundreds of investors using it daily. What really caught my attention isn't the "hundreds of users" number — it's the learning loop behind it.
When investors use PAT for real research, backend AI agents continuously scan these interactions, judge where PAT went wrong, generate human-audited benchmarks, and then drive updates to context and harnesses. In other words, the system doesn't just serve one user — it improves itself with every real use, ultimately benefiting everyone.
There's also an interesting button in the demo called Teach. If a user finds PAT's output needs correction, they can click Teach. A backend agent analyzes whether there are behavioral errors, context gaps, or user guidance that could have been provided earlier, then generates an expected-failure benchmark, iterates to fix it, and even generates a PR.
This goes further than "humans grading AI."
It turns human feedback into a system improvement task.
This is the key to AI agents truly entering production: it's not about having a model that answers questions — it's about having a mechanism that catches errors, reviews them, and transforms them into new context, new tests, and new processes.
Without this loop, more AI agents just mean more uncontrollable conversation windows for the organization.
With this loop, AI agents can become part of organizational capability.
Treating AI Agents Like Compilers
Santi Weight gave a powerful analogy in the second half.
He said his background is in compiler theory and programming language design, and they gradually realized that doing agentic coding should be like building compilers.
"Compilers compile user code into JavaScript. Coding agents compile user prompts or plans into Python."
This analogy is accurate.
If you treat AI writing code as "let the model freestyle," the system will be hard to stabilize. But if you treat it as a compiler problem, you naturally care about determinism, correctness, reliability, static analysis, dependencies, and verification.
Bridgewater's approach is to have the chat agent first form an analysis plan, then call coding agents. Coding agents don't randomly write a pile of code — they generate code based on tasks in the plan, then pass it through verification agents to check if it meets task requirements. If not, they continue correcting.
They even use static analysis and cache annotation to make repeated runs nearly instant.
This is why I think this talk isn't just a financial industry case. It describes a more universal enterprise AI agent architecture: turn natural language requirements into structured plans, turn plans into executable code, pass execution results to a verification system, and turn errors into continuously improvable feedback.
This pattern will enter many industries.
Consulting, manufacturing, pharma, legal, education, finance, cross-border e-commerce — all will face the same question: how to turn expert processes into executable, auditable, improvable AI work systems.
In Closing: Don't Just Buy Models — Distill Methods
This content doesn't constitute investment advice or discuss specific trading judgments. Bridgewater's value isn't in telling us where markets will go — it's in showing us how a mature organization connects AI agents into its knowledge system.
For Chinese entrepreneurs, the real takeaways from this case may have three layers.
First, enterprises need to rethink "knowledge management." In the past, much knowledge management was just document archiving — written and then forgotten. In the AI era, knowledge management must become working context that AI agents can read, invoke, and verify.
Second, enterprises need to rethink "processes." Processes aren't for people filling out forms — they're for making complex work decomposable, parallelizable, checkable, and reusable. Processes that can be clearly broken down are the ones AI agents can take over.
Third, enterprises need to rethink "feedback." Every correction an employee makes to AI shouldn't just be a chat log — it should become material for making the system better.
Bridgewater used fifty years of yellow paper rules to raise today's PAT.
What's most worth learning from this isn't the yellow paper, nor the finance — it's the organizational habit of writing things down over the long term, structuring methods, and turning errors into improvements.
AI will make many enterprises realize for the first time: what's truly scarce isn't models — it's work methods that have been seriously distilled.
Whoever writes them down first will be the first to have the soil for AI agents to truly work.
About MindsLeap
MindsLeap is an AI-native organization transformation accelerator.
We work closely with Founders Space, a Silicon Valley innovation incubator, to connect global AI frontier insights, Silicon Valley tech entrepreneurship ecosystems, and real transformation scenarios for Chinese entrepreneurs.
Around AI-native organization building, MindsLeap is building a transformation ecosystem for entrepreneurs, startup founders, AI engineers, industry experts, and investors — helping enterprises move AI from cognition, strategy, and tools into real organizational capabilities, business processes, product innovation, and growth systems.
This article was translated and adapted from the Chinese original with AI assistance.