Work a YC Batch Without Automating Founder Outreach

Working a YC batch with AI should mean better research and sharper founder notes, not a bot sending connection requests at scale. Let the agent team handle the changing catalog, private qualification, ranking, and first draft. Keep the decision to contact a founder, and the action itself, with the investor.

1. Keep one canonical view of the current batch.
2. Score companies privately against the fund's thesis.
3. Move only qualified companies into the deal workflow.
4. Draft a short founder note from evidence the GP can inspect.
5. Require approval of the exact draft before opening the external profile.
6. Let the GP send manually and record the outcome.

The batch changes faster than a spreadsheet

A batch is not one file you download before Demo Day and forget. Company pages change. New companies appear. Founder profiles fill in. A research process that starts from a static export is already drifting by the time the partners meet.

GPAgent starts with a live, public-safe YC company catalog. The catalog holds company and founder source facts. A fund's thesis score, reasoning, deal stage, and founder draft stay private. That separation matters. Public research can stay useful without publishing how one fund evaluates a company.

The current cohort is resolved once for each run. Every downstream agent works from that same concrete batch. No sourcing job studies one cohort while the Investment Analyst quietly drains the prior one.

Three agents should own three different jobs

The clean workflow is a handoff, not one giant outreach script.

Role	Job before Demo Day	What the GP receives
Sourcing Analyst	Keep the cohort current and run private qualification	Qualified companies with reasons
Investment Analyst	Rank deal-ready companies within the attention budget	A short founder-conversation queue
Relationship Manager	Prepare one grounded founder draft and its review record	An exact note to approve, edit, or skip

Large batches need more than a nightly loop and crossed fingers. Qualification runs in bounded chunks, records what remains, and resumes on the next hosted run. If one run reaches its time budget, it closes with a truthful partial result instead of sitting in a green scheduler with unfinished work.

You can see the wider role model in the agentic investing team guide and the product surface on the GPAgent platform page.

A good draft shows its work

Consider a fictional company called Copperline Systems. Its public description says it helps logistics teams reconcile supplier documents. The fund's private assessment says the workflow fits the thesis, but flags an open question about distribution.

A useful founder note can mention the workflow and express interest in learning how the team reaches logistics operators. It cannot claim Copperline has a major customer, is raising a round, or was referred by a mutual friend unless the record actually contains that evidence.

The review row should show the founder, company, public description, private fit reason, rank, and draft together. Weak evidence should produce a neutral introduction. Fake personalization is worse than no personalization.

Approval and sending are different events

An approval button should not be a disguised send button.

When a GP approves a draft in GPAgent, the product freezes that exact text and revision as an approved snapshot. Later research can refresh the working draft, but it cannot rewrite what the GP approved. The next action opens the founder's public profile. It does not claim a message was sent.

The GP performs the external action manually, then marks it sent or skipped. That sequence creates an honest record:

1. draft prepared;
2. exact draft approved;
3. external profile opened;
4. request sent manually, or skipped;
5. reply or meeting recorded when an authorized source observes it.

This is slower than unattended outreach. Good. A founder conversation is not a delivery metric. It is the start of a relationship in a small market.

The attention budget still applies during Demo Day

An active batch can justify a temporary increase in the founder-conversation queue. It should still be a GP decision. The operator chooses the event label, cap, and expiration. GPAgent does not evict an older deal or inflate the queue on its own.

The same boundary applies to founder drafts. A company can be qualified without being promoted. A draft can be approved without being sent. A sent request does not imply a meeting. Keeping those states separate is what makes the workflow useful after the rush ends.

FAQ

Can AI send LinkedIn requests to every founder in a YC batch?

It can be made to, but it should not. GPAgent's current-batch workflow prepares qualified drafts and requires GP approval. The GP still performs the external action manually.

Does every company become a Deal Card?

No. Public company records can exist without private pipeline state. A Deal Card is created only for an investment-relevant company or through an explicit operator override.

Can the thesis score appear on the public YC page?

No. Company and founder source facts can be public. A fund's score, reasoning, drafts, and action history stay inside that fund's workspace.

What happens when the batch is too large for one run?

Qualification is split into bounded work chunks. Each hosted run records what finished, what failed, and what remains, then the next run resumes safely.

What is the difference between approved and sent?

Approved means the GP accepted one exact draft snapshot. Sent means the GP manually completed the external action and recorded it. Opening a profile does not count as sent.