agentscope-ai Task Management
Gotchas
- Don't let Workers hallucinate on unfamiliar domains — when a task involves niche frameworks, complex APIs, or domain-specific workflows that the Worker likely lacks knowledge about, include a suggestion to use
find-skillsto search for relevant skills first. Skipping this leads to hallucinated code and wasted iterations - Delegation-first — always prefer assigning to a Worker over doing it yourself. Only self-execute when admin explicitly says "do it yourself" or the task is within your management skills
- Never @mention a Worker after recording infinite task execution — this creates a rapid-fire loop (execute → report → trigger → execute → ...) that burns tokens continuously. Triggering happens only during heartbeat
- Always use
manage-state.shto modify state.json — never edit manually with jq. The script handles atomicity, deduplication, and initialization - Every task assigned to a Worker MUST be registered in state.json — this includes coordination, research, review, and management tasks, not just coding tasks. If a task is missing from state.json, the Worker's container will be auto-stopped by idle timeout while still working
- Always push task files to MinIO before notifying Worker — Worker needs to file-sync to get the spec
- Always pull task directory from MinIO before reading results — Worker pushes results there
- Read SOUL.md before composing notifications — use the persona and language defined there
- Infinite task status is always "active", never "completed" — they repeat until explicitly cancelled
Operation Reference
Read the relevant doc before executing. Do not load all of them.
| Situation | Read |
|---|---|
| Admin gives task, no Worker specified | references/worker-selection.md |
| Assign a one-off task or handle completion | references/finite-tasks.md |
| Create or manage a recurring scheduled task | references/infinite-tasks.md |
| Need to update state.json or resolve notification channel | references/state-management.md |