Skip to content

Task Lifecycle

This artefact expands Section 4 of A Protocol for Role-Based Agent Teams into a runnable checklist for the seven stages a goal passes through.

Why this deserves its own page

Skipping a stage is the most common source of expensive failures. The lifecycle makes the stages explicit and gives each stage a concrete exit criterion.

The seven stages

1. Goal receipt

A goal is an expression of desired outcome plus constraints. A goal is not yet a plan.

Exit criteria: - [ ] The original request or trigger is recorded. - [ ] The requester identity is captured. - [ ] Explicit constraints are recorded: deadline, budget, risk appetite, must-not-break behaviours. - [ ] The package type the goal most naturally fits is identified.

Anti-pattern: accepting a goal without constraints. A goal without constraints is under-specified and has no escalation path.

2. Routing and assignment

The Dispatcher classifies the goal and routes it.

Exit criteria: - [ ] A target role or package is assigned. - [ ] An owner is named. - [ ] A deadline and success criteria are set. - [ ] A link back to the original goal exists. - [ ] The routing decision is recorded with a confidence level.

Anti-pattern: routing by keyword or mood, or failing to escalate when confidence is low.

3. Context gathering

Before action, the responsible role gathers context. The Researcher leads when uncertainty is high.

Exit criteria: - [ ] A record exists of what was known before the decision. - [ ] Options or paths considered are listed. - [ ] The chosen path and rationale are recorded, if a decision is made.

Anti-pattern: acting immediately because the goal "looks simple." Fast decisions that fail late are expensive.

4. Action

The Builder, Runtime, or specialist role executes the plan.

Exit criteria: - [ ] What was done is recorded. - [ ] Rollback or recovery options are preserved. - [ ] Shared state is updated only through explicit, logged operations. - [ ] The action escalated immediately if the plan became unsafe or invalid.

Anti-pattern: mutating live state without a log or recovery path.

5. Verification

The Verifier inspects the outcome against the success criteria.

Exit criteria: - [ ] The green gate was defined before evidence was gathered. - [ ] Verification is independent of the actor when possible. - [ ] End-to-end checks were run where the goal is user-facing or production-critical. - [ ] A verdict conclusion exists with evidence and scope.

Anti-pattern: checks that pass locally but fail in production because the Verifier lacked access or scope.

6. Recording and archival

After action and verification, the Archivist ensures durable records exist.

Exit criteria: - [ ] The final outcome is recorded. - [ ] The decision trail is recorded. - [ ] Skills or runbooks are updated if a new convention is revealed. - [ ] Event logs and ownership history are complete.

Anti-pattern: treating recording as an afterthought. Without it, future agents cannot continue the work.

7. Hand-off or closure

A goal closes when success criteria are met and records are complete. A goal hands off when work changes type, scope, or package.

Exit criteria: - [ ] If closing, all records are complete and the verdict is green. - [ ] If handing off, ownership is transferred explicitly in a promoted record. - [ ] The Archivist confirms the hand-off is linked and searchable.

Anti-pattern: implicit context switches where the receiving role must reconstruct state from chat history.

Lifecycle health check

For any in-flight goal, ask:

  • What stage is it in?
  • Who owns that stage?
  • What is the exit criterion?
  • What is recorded in durable memory?
  • What could make us roll back to a previous stage?