Airflow + Workflow demo
A workflow agent for apache/airflow — meta-poetic, since Airflow itself is a workflow engine.
Reproduce
cd /tmp
git clone https://github.com/apache/airflow.git
cd airflow
git checkout <pinned-SHA>
uvx harnessforge init --no-llm --blueprint workflow-agent
What gets written
AGENTS.mdwith the workflow loop (decompose → call → check → log → next)SOUL.mdsetting the "operational, low-drama" voice — appropriate for an infrastructure projectTOOLS.mdrecommendingfilesystem,fetch,shell,github,postgres+ the tool-call logging contractMEMORY.mddefiningtool_log+ skill-derived heuristicsSKILLS/decompose-task/— plan generator with idempotency markersSKILLS/call-tool-with-retry/— exponential-backoff retrySKILLS/check-result/— independent verification of expected effects- Adapter files for Claude Code, Cursor, Continue, Windsurf, Codex
Why this demo
- High recognition value — every data engineer knows Airflow
- Correctly recommends Airflow-aware MCPs through the catalog
- The
idempotentvalidator enforces the same property Airflow itself depends on for retries - Demonstrates that "workflow" is not just a fallback — it's a real production pattern
Validators
After init:
$ harness verify --json
{
"blueprint": "workflow-agent",
"checks": [
{"name": "structure", "status": "pass"},
{"name": "tool_log", "status": "skipped"},
{"name": "idempotent", "status": "skipped"}
],
"summary": {"total": 3, "passed": 1, "failed": 0}
}
The tool_log and idempotent checks skip cleanly until an agent actually produces a /tmp/airflow-workflow-output.json. Once your coding agent runs a workflow, the validators score it against the contract.