AgentRC Agentic Workflow Conventions
September 20, 2026 ยท View on GitHub
- Treat
.github/workflows/*.mdas source and.lock.ymlplusagentics-maintenance.ymlas generated artifacts. - Never edit generated workflow files by hand or update their action pins with Dependabot.
- Compile every workflow with the repository's pinned stable
gh-awversion. - Use
copilot-requests: writefor Copilot inference when organization policy permits it. Do not introduce a PAT unless a workflow requires cross-repository access or recursive CI triggering. - Keep the agent job read-only. Use safe outputs for GitHub writes.
- Agent-created code changes must be draft pull requests. Do not enable auto-merge in customer-facing examples.
- Every scheduled workflow must define a stop date, concurrency policy, per-run AI-credit cap, daily AI-credit cap, turn cap, and bounded output.
- Prefer deterministic setup and analysis before the agent starts. Give the model a small evidence file instead of asking it to rediscover facts.
- Use
/tmp/gh-aw/agent/for temporary evidence that should be included in run artifacts. - Read
AGENTS.mdfor repository architecture. Do not duplicate architecture details in workflow prompts. - Keep demo workflows manually triggerable and maintain a known successful run as a fallback for live presentations.