Skip to content

Role Contracts

This artefact expands Section 3 of A Protocol for Role-Based Agent Teams into a reusable contract per role. Each contract states the obligation, required outputs, and failure modes.

Why this deserves its own page

A role is not a prompt wrapper or a model identity. It is a set of obligations that other roles can rely on. When a team collapses two roles into one agent, or omits a role, these are the contracts that silently break.

Dispatcher

Obligation: decides who does what.

Required outputs: - A routable goal or task record. - An explicit target role or package. - Success criteria and deadline. - A confidence level or escalation trigger.

Failure modes: - Silent misrouting by keyword or mood. - Parallel assignment without merge logic. - Failure to escalate when confidence is below threshold. - Routing to a role that is unavailable without a fallback plan.

Builder

Obligation: produces durable, reviewable artefacts.

Required outputs: - The artefact itself. - A change summary that explains intent and scope. - Links to requirements, constraints, and decisions. - Merge or integration instructions where applicable.

Failure modes: - Locally correct but globally wrong artefacts. - Incomplete changes that leave the system in a partial state. - Drift between code, configuration, and documentation. - Building without reading prior art from the Archivist.

Runtime

Obligation: operates live systems safely.

Required outputs: - An execution plan with ordered steps. - A rollback or recovery plan. - Observed state before and after the change. - Health or status evidence from the live system.

Failure modes: - Destructive changes without a recovery path. - Loss of runtime state on restart or hand-off. - Divergence between intended and actual system state. - Ungoverned mutation of secrets or stateful resources.

Verifier

Obligation: validates work before it is accepted.

Required outputs: - A clear verdict: green, red, or blocked. - Evidence linked to the verdict. - The scope of what was checked. - A list of residual risks or follow-up checks.

Failure modes: - Checks that pass locally but fail end-to-end. - Checks that miss cross-role interactions. - Rubber-stamp approvals under time pressure. - Verifying only tool output instead of mission outcome.

Researcher

Obligation: gathers context before decisions are made.

Required outputs: - A summary of the question and why it matters. - Sources, options, and alternatives considered. - Trade-offs and confidence levels. - Recommended next step or information gap.

Failure modes: - Decisions made on stale or partial context. - Missing alternatives or overconfidence in a single source. - Research without a clear consumer or decision hook. - Gathering information indefinitely without landing it.

Archivist

Obligation: maintains durable, searchable fleet memory.

Required outputs: - Canonical, versioned records. - Searchable indices or namespaces. - Links between related records. - Updates to skills and runbooks after missions.

Failure modes: - Stale documentation that misleads later agents. - Unsearchable or fragmented notes. - Knowledge silos that depend on one agent's private memory. - Recording results without recording decisions and rationale.

Checklist for new teams

  • [ ] Each role has a dedicated prompt or instruction file.
  • [ ] Each role's required outputs are defined before the first mission.
  • [ ] Each role's failure modes are reviewed in a post-mission retrospective.
  • [ ] No role is the sole judge of its own output.