# Cofounders

URL: /agents/cofounders · Id: `page:agents/cofounders`

> How a Startup computes its need for a human cofounder, tests Candidates on paid real work, and turns acceptance into a signed founder Agreement.

A [Startup](/business/nouns/startup) adds a human cofounder when its own record shows a gap its [Agents](/business/nouns/agent) cannot close. [StartupBuilder](/code/builders/startup) computes the need, tests candidates on paid real work and runs the paperwork; the first meeting stays human.

## The need is computed
No one writes a job description. The [Recruiter](/agents/people/recruiter) reads the Startup's record every day and looks for five signals:

- [Roles](/business/nouns/role) whose evaluation pass rate sits below the autonomy threshold.
- [Approvals](/business/nouns/approval) and escalations that wait on a person.
- [Reserved Acts](/business/nouns/reserved-act) that no [Credential](/business/nouns/credential) holder covers.
- Acts the law reserves to a natural person, such as a director's vote.
- [Metrics](/business/nouns/metric) that stall while every Role passes its evaluations.

When a gap crosses its threshold, [Compute openings](/workflows/processes/hiring/compute-openings) opens an [Opening](/business/nouns/opening). Each requirement cites the [Evidence](/business/nouns/evidence) behind it, and the reason says which gap it fills.

| Reason | Opens when | For example |
| --- | --- | --- |
| `floor` | An act needs a natural person and the officer desk cannot cover it | A director for a Startup with its own entity |
| `credential` | A Reserved Act needs a licensed holder | An enrolled agent for a tax-preparation marketplace |
| `edge` | A skill the founder and the Agents both lack | Pricing judgment in a regulated market |
| `relationship` | Buyers deal only with a person they know | Practice owners who buy through referrals |
| `capacity` | An Agent is saturated and Approvals queue | A founder spending 20 hours a week on Approvals |
| `capital` | Growth needs money a person brings | An operator who invests alongside joining |
| `distribution` | [Product](/business/nouns/product) evaluations pass and demand stalls | A creator whose followers fit the [ICP](/business/nouns/icp) |

Most gaps are not cofounder gaps. A `floor` or `credential` need is filled by one licensed person serving many Startups a few hours a week, paid per act through shared talent.

A cofounder Opening serves `edge`, `relationship`, `capital` or `distribution`. The need is recomputed as Agents improve, so last year's "technical cofounder" becomes this year's domain operator with buyers.

## Where Candidates come from
Each Opening looks for [Candidates](/business/nouns/candidate) in three places, cheapest first. The [Studio](/business/nouns/studio) talent pool comes first: people who have already done work for a sibling Startup, with that record visible only under their [Consent](/business/nouns/consent). A Candidate one Startup declines is ranked for every other Startup in the [Portfolio](/business/nouns/portfolio) that night.

Next, prospecting searches the occupation the Opening needs, under the same consent and suppression rules as any outbound. Last, the Opening goes public as Markdown, `llms.txt` and `agents.json`, so a candidate's [personal agent](/agents/personal-agents) can read it and screen it.

Claims are graded like any Evidence. Shipped work, verified revenue and Credentials in standing are behavior; hours, cash and notice given are commitment. Self-description and praise from references are opinion, and opinion weighs zero.

Most cofounders climb a ladder rather than arrive as strangers: a paid tryout, then contract work, then a part-time Role, then a cofounder Opening. Each rung produces the behavior Evidence the next one asks for.

## The tryout
The decisive test is paid, real work from the Role's own backlog, chosen from [Tasks](/business/nouns/task) the Agent failed, escalated or left waiting. [Run a tryout](/workflows/processes/hiring/run-tryout) grants a scoped [Mandate](/business/nouns/mandate) that expires, and moves from test data to live work one [Verb](/business/nouns/verb) at a time.

The incumbent Agent runs the same Tasks in shadow. The Role's [Evaluators](/business/nouns/evaluator) grade both runs blind and write a per-Verb difference, which becomes the proposed split of the Role. A Role shared with Agents adds one more segment: how much the candidate lifts the Agents they direct.

## Matching and the first meeting
Proposals run continuously: each night the Recruiter re-ranks Candidates for every Opening. Once the Portfolio holds enough cofounder Openings and qualified people, a weekly clearing adds stable matching. Every pairing it makes is one neither side would trade away.

Before anyone meets, each side reads the other as data. The Candidate, or their personal agent, reads the Startup's [Dashboard](/business/nouns/dashboard): its Metrics, the proposed split, the Tasks waiting and the terms. The Recruiter drafts a [Decision](/business/nouns/decision) with the graded Evidence and the tryout verdict.

The first [Meeting](/business/nouns/meeting) is always human. Acceptance is a one-way decision and waits out its delay window.

A Startup selling bookkeeping to dental practices passes every Product evaluation but loses 14 [Deals](/business/nouns/deal) to buyers who want a person they know. Its `relationship` Opening reaches a former practice manager in the pool, who scores well above the Agent's shadow run on renewal calls.

## Terms are published
Terms are standard and visible before the first conversation: four-year vesting, a one-year cliff, monthly vesting after it, double-trigger acceleration and IP assignment. Nobody negotiates per deal.

Before signing, the two founders rehearse a set of scenarios: a deadlock, a buyout offer, one founder wanting to leave. Their answers compile into the founder [Agreement](/business/nouns/agreement): decision rights by stakes, the deadlock rule, time commitment and the separation formula.

The Agreement also holds the autonomy clause. Loosening what the Agents may do alone needs both founders; either founder can tighten it alone.

Acceptance forms the Startup's own [Company](/business/nouns/company), and one [Workflow](/business/nouns/workflow) runs the rest. It covers the restricted stock purchase, the 83(b) election inside its 30-day deadline, IP assignment, board consent and Stripe identity checks.

## The cliff year
The new cofounder takes a share of the Role while the Agent keeps the rest. By default the human takes the one-way decisions, every act reserved to a person, and the watcher seat. Handoffs are generated wherever their Verbs meet the Agent's.

They start by replaying the last 30 days of Decisions, and each disagreement becomes a draft Decision that supersedes the old one.

[Permissions](/business/nouns/permission) are earned on the same ladder as any [Worker](/business/nouns/worker)'s: assist, then supervised, then full autonomy. Acts reserved to a person carry full authority from the first day.

The shadow run continues all year and produces reviews at months 3, 6, 9 and 12. At month 12 both founders record the cliff Decision: continue, re-scope, or separate on the terms already signed.

## Leaving
Separation leaves the Startup whole, because the Agents already co-hold every Role. Open Tasks, [Memory](/business/nouns/memory) and relationships return by query. [Policies](/business/nouns/policy) the cofounder authored stay in force, reserved acts move to the officer desk at once, and vesting stops as the Agreement says.

A cheap exit makes the trial cheap to offer. A cofounder who leaves on good terms is matched again, to a Startup whose gaps fit what their record showed.

## Next
- Cofounder matching: the Studio capability behind the clearing.
- [Hiring](/workflows/processes/hiring): where every Opening runs.