An investment thesis sat in a static memo, so nobody could see which of its claims still held as deals changed.
Break the thesis into linked claims, risks and initiatives, and design an agent that reasons across every deal the fund holds.
Thousands of initiatives proposed and amended through it.
- Initiatives proposed and amended
- Thousands
- In the agent’s reasoning scope
- Every deal
- Linked types: claims, risks, initiatives
- 3
Context
- Company
- VectorShift (YC S23)
- My role
- Built the product and designed its agentic layer
- Users
- Private equity deal teams and portfolio operations
- Scope
- Every deal the fund holds
The problem
A fund writes an investment thesis when it decides to back a company. After that, the thesis usually lives in a memo.
The claims inside it change status quietly. This market grows. This margin expands. This team can execute. The risks that challenge those claims show up in board updates and diligence notes, not in the memo.
So when a partner asks whether to push an initiative or pull back, the reasoning gets rebuilt from memory and scattered documents, one deal at a time.
- Theses were read at underwriting and at exit, rarely in between.
- The same kind of risk appeared across several deals without anyone connecting them.
Constraints & non-goals
Constraints
- The agent has to reason across every deal the fund holds, not one document at a time.
- Investment decisions stay with people. The agent makes the argument. It does not take the action.
- Every argument has to point back to the claims, risks and sources it depends on.
Out of scope
- Automated investment decisions.
- Market forecasting. The agent reasons over the fund’s own information.
- Replacing the investment committee memo.
Options I weighed
- NOT CHOSEN
Chat over the memo folder
Fluent answers with no structure underneath. Retrieval finds passages. It does not know that a claim in one deal is undermined by a risk in another.
- NOT CHOSEN
A manual thesis tracker with no agent
Structure without leverage. At fund scale, nobody keeps the links up to date by hand.
- CHOSEN
Linked thesis objects plus an agentic layer
The structure gives the agent something real to reason over. The agent keeps the structure useful by proposing and amending initiatives.
The spec
Users
- Partners and principals who own a thesis.
- Portfolio operations teams who run value-creation initiatives.
Requirements
- A thesis breaks down into claims, risks and initiatives that link to each other.
- The agent argues for or against a move using information from every deal in the fund.
- People can propose and amend initiatives.
- Every argument cites the claims, risks and sources it relies on.
Acceptance criteria
- When a new risk is added to one deal, the agent surfaces other deals with linked claims.
- Every argument shows the evidence for and against.
- A person accepts, edits or rejects every proposed change.
Guardrails
- The agent never changes a thesis or initiative without a person accepting it.
- Arguments without cited evidence are not shown.
What shipped
Decompose
The thesis is broken into claims, risks and initiatives.
Link
Objects connect within a deal and across deals.
Reason
The agent builds the case for and against a move across the whole fund.
Decide
People propose and amend initiatives.
What changed
Thousands of initiatives have been proposed and amended through Deal Intelligence.
- The thesis became something teams use between underwriting and exit, not a memo read once.
- It is the clearest example of how I design agent systems: structure first, then reasoning, then a human decision.
HONEST CAVEATThis measures workflow adoption. I don’t claim an effect on investment returns.
Reflection
The design question that mattered most was not what the agent could say. It was what the agent was allowed to change.
If I did it again, I would put the linked structure in front of users before the agent, so the objects the agent reasons over are already familiar when its arguments show up.
Last updated · Company tenure is VectorShift · Jul 2025–present, not a claim about individual project dates.