Skip to content

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.md with the workflow loop (decompose → call → check → log → next)
  • SOUL.md setting the "operational, low-drama" voice — appropriate for an infrastructure project
  • TOOLS.md recommending filesystem, fetch, shell, github, postgres + the tool-call logging contract
  • MEMORY.md defining tool_log + skill-derived heuristics
  • SKILLS/decompose-task/ — plan generator with idempotency markers
  • SKILLS/call-tool-with-retry/ — exponential-backoff retry
  • SKILLS/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 idempotent validator 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.