| name | execution-mode-routing |
| description | Choose how new Prometheus work should execute: inline, planned, ephemeral background, durable task, delegated subagent, coordinated team, or scheduled job. Use when execution mode, persistence, parallelism, ownership, or restart survival is unclear; do not use merely to inspect or control work already running. |
Execution mode routing
Choose the smallest execution surface that safely reaches the requested outcome.
Decision order
- Inline: Use for short work that can complete in the current turn without durable monitoring.
- Interactive plan: Use when several dependent steps benefit from visible progress but the current agent still owns execution.
- Ephemeral background agent: Use for bounded asynchronous work that may outlive the turn but does not require restart durability.
- Durable task: Use when work must survive restart, expose persistent state, or support later control and verification.
- Delegated subagent: Use for one independent specialist batch with a clear boundary and deliverable.
- Coordinated team: Use when multiple independent batches can run concurrently and their outputs require synthesis.
- Scheduled job: Use when execution is tied to a future time or recurrence rather than starting now.
Selection questions
- Can the work finish safely in the current turn?
- Must it survive a restart or disconnection?
- Does it need later pause, resume, cancellation, or retry?
- Are there genuinely independent workstreams?
- Does another agent need distinct tools or context?
- Is a future or recurring time part of the requirement?
- What artifact or external state proves completion?
Do not use backgrounding merely to avoid waiting. Do not create a durable task for trivial work. Do not spawn multiple agents for tightly coupled edits to the same files. Do not schedule work that should run immediately.
Handoff contract
For delegated, durable, team, or scheduled work, provide a self-contained objective, scope boundaries, relevant paths/context, prohibited actions, validation requirements, expected artifacts, and completion evidence. Preserve ownership: the parent remains responsible for integrating and verifying child output.
After starting
Return to task-lifecycle for inspecting, controlling, recovering, or verifying existing work. Use scheduler-operations-playbook for schedule-specific mutation and delivery diagnosis.
Read task-lifecycle/references/detailed-guide.md only when older tool-by-tool execution guidance is needed.