SKIP TO CONTENT
AI AGENTS5 minAgents in Production / Ep. 2 · Stage 2-1

11 AI Agents Inside a $3B Fund: The Base and the Guardrails

One question could mean opening five systems to find one document among 1.4 million files. What did 11 AI agents share to answer it with a source?

XIN
Pixel-art illustration of eleven identical agent modules arranged in a ring around one shared glowing foundation slab they all connect to.

IN 30 SECONDS

An AI agent is software that uses a language model to do one defined job, like finding a document. I built 11 of them for a US real-estate fund managing roughly $3B in assets. This stage shows the foundation they shared and its checks.

KEY TAKEAWAYS · 3

  • I shipped 11 production AI agents for a US real-estate private-equity fund with roughly $3B in assets under management.
  • All of them stood on one foundation: about 2 TB and 1.4M files from 13 years and 13 sources, in one searchable store.
  • The guardrails were the real engineering: every claim cites a source, every number is checked, and thin evidence goes to a person.
On this page

One engineer, 11 agents, one fund

IN SHORT The hard part was the guardrails all 11 agents shared, more than the agents themselves.

I worked remotely as the An engineer who works inside the client’s own workflows. AI engineer for a US real-estate private-equity fund. It manages roughly $3B in assets.

The platform had 11 AI agent Software that uses a language model (an AI that writes text) to do one defined job, within limits people set. See the glossary at first delivery and 25 modules at handover. This stage covers that first architecture. All 11 stood on one shared foundation, though each new agent still needed its own tests and permissions.

TWO WAYS TO GROW A PLATFORM

11 separate projects

Each agent gets its own search, tests and safety checks. Securing agent number 11 means starting over.

One shared foundation

One search layer, one test suite, one set of guardrails. Agent 11 inherits them, then adds its own tests and permissions.
The road not taken, and the one I took.

Read the case study

Coming up next: So what exactly counts as an agent here?

Words to know

IN SHORT An agent here is narrow: one job, the fund’s own documents, and a source for every answer.

WORDS TO KNOW

AI agent
Software that uses a language model (an AI that writes text) to do one defined job, within limits people set.
Retrieval
Searching a store of documents and pulling out the few that answer one question.
Model router
The component that sends each task to a cheaper or a stronger model, under a spending cap.

ADAM says: Picture a library where every desk shares one catalogue and each answer carries a shelf mark. Where it breaks: a librarian can answer from memory; these agents were designed not to.

Coming up next: But where did those documents live before?

The first problem was the data, not the AI

IN SHORT Before any agent was useful, the fund’s documents had to live in one place.

The records sat in 13 sources, built up over 13 years, in formats that didn’t agree. In all, roughly 2 TB.

1.4M

files, 13 sources, 13 years

Reported from my engagement, March to August 2026.

Answering one question could mean opening five systems to find the one document that held the answer. The hunting was the bottleneck.

So my first build wasn’t clever: one store you can query. “Where is the document that explains this?” went from an afternoon of digging to seconds. Every agent’s Retrieval Searching a store of documents and pulling out the few that answer one question. See the glossary runs against it.

THE DATA FUNNEL
Flow diagram: 2 stepsFlow diagram: 2 steps. 13 sources, then 1 searchable store.113 sourcesStatements, filings,contracts, valuations, emails,built up over 13 years.21 searchable store~2 TB, 1.4M files, one placeevery agent asks.
13 sources built up over 13 years, funneled into the one store every agent asks.

ITEM GET

One store before any agent

Every agent’s retrieval runs against one searchable store, so the data work came first.

Coming up next: So how does each agent ask that store for evidence?

The shared stack under all 11 agents

IN SHORT Every agent asks for evidence the same way, through one shared stack.

FastAPI, the web layer, takes requests. Postgres, a database, holds the documents, and its pgvector add-on searches them by meaning.

A Model router The component that sends each task to a cheaper or a stronger model, under a spending cap. See the glossary picks a cheaper or a stronger model for each task. Its hard cost cap keeps 11 agents from quietly running up a bill nobody approved.

STEP BY STEP

Layered stack diagram: 6 layers, top to bottomLayered stack diagram: 6 layers, top to bottom. Layer 1, ENTRY: Web layer. Layer 2, SEARCH: Searchable store, ~2 TB of documents. Layer 3, ROUTER: Model router, Hard cost cap. Layer 4, MODELS: Cheap model, Strong model. Layer 5, CHECK: Does each claim match its source?, Do the numbers match?. Layer 6, FALLBACK: Abstain, Hand to a person.ENTRYWeb layerSEARCHSearchable store~2 TB of documentsROUTERModel routerHard cost capMODELSCheap modelStrong modelCHECKDoes each claim match itssource?Do the numbers match?FALLBACKAbstainHand to a person
The first check uses NLI (natural language inference): does the claim follow from the text?

Step 1. Arrive. A request comes in through the web layer.

Pick a step to see it move
  1. Arrive A request comes in through the web layer.
  2. Find the evidence The agent searches ~2 TB of documents. No document, no answer.
  3. Pick a model The router picks a model within the cost cap.
  4. Draft The model is meant to draft from the documents found, not from memory.
  5. Check Each claim is tested against its source, and so is each number.
  6. Fall back Thin evidence: the agent abstains and a person takes the case.
One request through the shared stack: evidence first, checks before any answer.

IN PLAIN WORDS

The agents were designed to find the documents first, answer only from them and check the answer, never to answer from memory. Thin evidence goes to a person. All of it was tested and red-teamed before launch.

STAGE CLEAR

What you now know

  • All 11 agents shared one searchable store, one stack and one set of checks.
  • Retrieval comes first: agents are designed to answer from what they found, not from memory.
  • A model router picks a cheaper or stronger model, under a hard cost cap.

Check yourself

Tap or click a question to see the answer.

A new agent joins the platform. What does it get for free, and what does it still need?

It inherits the store, the stack and the guardrails. It still needs its own tests and permissions.

Next stage:How does the safety layer decide between answering and saying “I don’t know”?5 min
Adam Boudjemaa
Adam Boudjemaa

Web3 x AI engineer, former CTO of Integra. Named author (1 of 5) of ERC-3643, first author of ERC-6960, co-author of ERC-7410, and co-author of ERC-8203, which is still a draft.

Have a project to ship or a role to fill?

Discuss your project or role

My work at the web3 x AI intersectionOpen-source tools I build

LinkedInTwitterGitHub
Follow the engineering notes

Get new articles on production AI and smart contracts by email.