draft-job-contract

Draft a job contract

A job contract is the written job an AI agent is employed to do: what it owns, what one unit of work is, what counts as acceptable, what it may spend, and what the same work costs when a person does it. Metrx scores the agent against this contract.

Steps

1. Collect from the person, asking only for what they have not said:
   • template: one of support_triage, sdr_qualification, internal_assistant, coding_patch.
   • owner: the accountable person (name or email). Required.
   • Optional: agent_key, workload, objective, unit_of_work, budget_usd with budget_period (daily, weekly, monthly), human_baseline (cost_per_unit_usd, optional minutes_per_unit, and basis, which says where the number comes from), notes. Do not guess the human baseline. If the person has no number, leave human_baseline out; the review will show "insufficient data" for the comparison, which is the honest state.
2. Call metrx_draft_job_contract with those fields. No key is needed; the tool reads nothing and stores nothing.
3. Relay the result:
   • The draft as a readable contract: objective, unit of work, acceptance criteria, budget, human baseline (each owner-declared; the result's label is proposal-not-measured, say so).
   • errors: each one blocks publishing; say which field and why.
   • warnings: each one is a gap the person should know about (for example a missing baseline). Relay them; do not resolve them silently.
4. If the person wants to publish it, say that publishing uses the metrx_set_job_contract tool (the result's publish block names the path and scope), which needs a Metrx API key with the admin scope (Settings → Security), and that the contract's owner becomes the agent's owner on the review. Only call metrx_set_job_contract when they say so.

Rules

• Every number in the contract is the person's own input and is labelled owner-declared. Do not present a template default as a measurement.
• Do not promise what the contract will achieve. It is the yardstick, not the result.