Recruiting
How a Startup computes its need for a Human from its own record, tests Candidates on paid real work against its Agents, and hires on Evidence.
No one at a Startup writes a job description. The Recruiter computes each need for a Human from the Startup's record, tests Candidates on paid work against the incumbent Agent, and hires on Evidence.
The need is computed
Every Role starts with an Agent in it. A Human is worth adding only where the record shows a gap the Agents cannot close. Compute openings reads five signals every day:
- Roles whose eval pass rate sits below the autonomy threshold for a decision class.
- Escalations and Approvals that waited on a person.
- Legal floor acts with no natural person to perform them.
- Reserved Acts that no Credential holder in the Startup covers.
- Metrics that stall while the Agents work.
When a gap crosses its Setting, the Workflow opens an Opening, unlisted at first. Every requirement on it cites the Evidence that produced it. The need is recomputed as Agents improve, so an Opening for a builder last quarter becomes an Opening for a seller this quarter.
Kinds and reasons
An Opening states what kind of commitment it asks for and why the Startup needs a person at all.
| Kind | What the Human does | Paid by |
|---|---|---|
floor | Performs the acts the law reserves for a person, across one or many Startups | Cash per act |
contractor | Takes a bounded share of a Role for a fixed scope | Cash, under an engagement Agreement |
employee | Takes a standing share of a Role | Salary and an option grant |
cofounder | Owns a share of the Startup's judgment and its outcome | Equity on published terms |
The reason is one of seven: floor, credential, edge, relationship, capital, capacity, or distribution.
Take an autonomous Startup that prepares visa petitions for employers. Agents gather documents and draft every petition, but the filing is a Reserved Act for a licensed attorney. Compute openings opens a floor Opening with reason credential, a few hours a week.
A headless scheduling API for veterinary clinics shows a different gap. Its technical founder ships quickly, but deals with clinic groups stall at the second call while the Sales Lead Agent wins single clinics. The Opening is cofounder, reason relationship: someone clinic owners already trust.
The match is on the residual, never on similarity. A Candidate is scored on what the Startup needs, minus what the founder covers, minus what the Agents already do.
Sourcing in order
Each Opening looks for people in three places, cheapest first.
- A Startup in a Studio's Portfolio queries the Studio's talent pool, where each Person's record arrives with past Evidence and their Consent.
- Prospecting with an occupational ICP searches for people who do the work today, under the same consent and suppression rules as any outbound Message.
- The Opening is listed publicly, in a form a candidate's personal agent can read and answer.
The Best users Play adds a fourth source: Customers whose own results show they already understand the work.
Screening reads Evidence
An Agent screens each Candidate in writing, over email or chat, on the Candidate's schedule. It asks only what the Opening's Hypothesis still lacks, so a strong record means a short screen.
Every claim is graded. Behavior counts most: shipped work, revenue verified at its source, Credentials in good standing, and acts performed.
Commitment counts next: hours offered, notice given, money put in. Self-description and praise with no incident behind it weigh zero.
Reference checks ask what the person did in a named situation, such as the week a launch slipped. Opinions of character weigh zero.
The paid tryout
The decisive test is real work. Run a tryout cuts Tasks from the Role's backlog, preferring those the Agent failed, escalated, or left waiting. The Candidate works under a scoped Mandate that expires with the tryout, and is paid in full whatever the grade.
The incumbent Agent runs the same Tasks in shadow. The Role's own Evaluators grade both arms blind. The result is a difference per Verb: where the Human beats the Agent, where they tie, and where the Agent is better.
That difference writes the hire's share of the Role. The veterinary Candidate closes three of the five clinic groups the Agent stalled on, and ties on everything else. The proposed split gives them group deals and leaves single clinics to the Agent.
For Roles shared with Agents, the tryout also measures how much the Candidate lifts the Agents they direct.
The Work trial Play is the default tryout: two weeks on the Startup's own Issues.
Deciding and agreeing
The Recruiter drafts a hiring Decision: the Hypothesis, the graded Evidence, the tryout result, and the proposed split. The Candidate, or their personal agent, reads the Startup's live view: its Metrics, the split, and the Tasks waiting on day one.
The first Meeting is between people. Hiring is a one-way decision, so the accepted Decision waits its delay window before anything is signed.
Terms are standard and published before anyone meets, so nothing is negotiated deal by deal. A cofounder joins on four-year vesting with a one-year cliff and IP assignment. Acceptance runs the paperwork: the Agreement, a Company where the Startup has none, the stock purchase, the tax election by its deadline, and identity checks.
Recruiting learns
Outcomes land in the same record: who vested at the cliff, how the Role's Metrics moved by month twelve, and who left. Recruiting runs as an Experiment on which Evidence and which tryout designs predict a good hire. Winning designs become Startup Template defaults.
A Candidate declined at one Startup is ranked for every other open Opening in the Portfolio that night. Their record travels only with their Consent.
Next
- Joining: the first weeks after the Agreement is signed.
- Sharing a Role: how the split divides the work.
- Cofounders and capital: recruiting across a Portfolio.