Agent Contract
Practicum for Volume 2, Chapter 3, "A Team with Agents." Version: 2026-06. Time: 15-20 minutes per agent. Use one sheet for one agent. Copy the template and fill it in for your case.
An agent is not a "magic employee." It is a role with clear boundaries, review, and a human owner. The contract turns "let's add AI" into an agreement the team can understand and enforce.
A minimum governance packet is the small set of documents that defines control and ownership. In that packet, the contract follows the Human Review Matrix. The matrix decides which tasks can be handed over. The contract defines one agent, its permissions, and its brakes. Do not scale right away. Test the agent on an evaluation set and in shadow mode first, where it drafts or observes without taking live action.
Template: 10 fields
| Field | What to enter |
|---|---|
| Role | the agent's name and function |
| Mission | one sentence explaining why it exists |
| Inputs | the data and formats it accepts |
| Tools | the systems, APIs, and files it may access |
| Memory | what it remembers between sessions, and what must never enter memory |
| Boundaries | what the agent never does without approval |
| Escalation | the conditions that send the task to a person |
| Review | how and when a person checks the result, rather than one AI checking another |
| Metrics | 1-2 numbers that show whether it works |
| Owner | the name of the person accountable for it |
Default rule: without Boundaries + Escalation + Owner, this is a wish, not a contract. All three are required from day one.
A measurable threshold gate: KPI -> escalation -> stop
A contract metric is not something to report at the end. It is a guardrail, a preset threshold that makes the agent stop and call a person. Connect three fields in one rule:
| What you define | Example |
|---|---|
| Signal metric | share of answers accepted without edits; unusual numbers; escalation rate; daily cost |
| Threshold, set in advance | "below 85% accepted without edits"; "difference > 10%"; "> N requests in a row"; "above the ₽/day limit" |
| Action at the threshold | pause and human review; return to manual work; owner-triggered shutdown |
Give the agent the smallest set of permissions it needs. This is called least privilege. A measurable threshold turns Metrics and Escalation into a working brake. Set the threshold before launch, not after an incident.
Example 1: first-line support agent
- Role: sorts incoming requests and answers common questions. Mission: remove routine work from the support team. Inputs: request text and answer base. Tools: read-only access to the knowledge base.
- Boundaries: does not promise compensation or touch billing. Escalation: complaints, refunds, or conflict go to a person. Review: a person audits 10 conversations a day. Metrics: share resolved without a person and number of escalations. Owner: head of support.
Example 2: reporting analyst agent
- Role: prepares the weekly sales summary. Inputs: CRM exports. Tools: read-only data access. Boundaries: does not send the report. Escalation: an anomaly > X% is flagged for a person. Review: the owner checks the totals against the source before sending. Metrics: time to finished report and share of figures accepted without edits. Owner: team analyst.
Pre-flight check
- The three required fields are complete: Boundaries, Escalation, and Owner.
- Access follows least privilege and covers only what the mission requires.
- Memory contains no secrets or personal data that do not belong there.
- A human review method is built into the flow.
- The team has agreed on the signal that shuts the agent down.
- A measurable threshold gate is set: metric + number + action (pause / rollback / stop).
Next
- Build shared context for the team and its agents with the team-brain starter kit.
- Decide what the agent may do and what a person should keep with the Human Review Matrix (Chapter 2).
- Review third-party skills before they enter your system: how to check them.
- Follow the full route in the Volume 2 Workbook.
Version: 2026-07-09. What changed: the contract now sits inside the minimum governance packet. An earlier update added the measurable threshold gate (KPI -> escalation -> stop) and pre-flight item. Do not put vendor names in the contract. Tools become outdated; the contract does not.