Workflow agent
Orchestrate tools + APIs to complete multi-step tasks. The generic default blueprint — use this when none of the more specialized ones fit cleanly.
When to use
You're building a general multi-step automation: ingest something, call APIs, write somewhere, notify someone. Think: Zapier-with-judgment, or a custom internal tool that needs to be reliable.
The loop
What gets generated
| File | Role |
|---|---|
AGENTS.md |
Workflow loop + conventions + forbidden lists |
SOUL.md |
"Operational, low-drama execution layer" voice |
TOOLS.md |
Per-tool docs + tool-call logging contract |
MEMORY.md |
Per-task tool log + skill-derived heuristics |
SKILLS/decompose-task/ |
Plan generator with idempotency markers |
SKILLS/call-tool-with-retry/ |
Exponential-backoff + idempotency check |
SKILLS/check-result/ |
Independent observation of expected effect |
scripts/test_task.sh |
Sample task replay |
scripts/verify_output.py |
Plan + tool log + idempotency check |
Validators
| Validator | What it checks |
|---|---|
structure |
Every blueprint-generated file present and parses |
tool_log |
Every plan step that uses a tool has a tool_log entry |
idempotent |
Final plan step is marked idempotent: true (safe to retry) |
Recommended MCP servers
filesystem— read/write project filesfetch— HTTP requestsshell— run commands (refuses forbidden_commands)github— PRs / issues / actionspostgres— read-only DB queries
Memory schemas
conversation— run_id + task + plan + tool_log + final_status + idempotent_marker
Hero demo
Airflow + workflow cookbook shows this blueprint on apache/airflow.