GitHub Agentic Workflows

Intelligent automation for GitHub. Run coding agents with strong guardrails and cost controls in GitHub Actions. Public Preview — may change significantly.

Key Features

  • Simple Markdown Files: Plain markdown instead of complex YAML
  • AI-Powered Decision Making: Context-aware, adaptive workflows
  • Multiple AI Engines: Copilot CLI, Claude Code, Codex, Gemini, custom processors
  • GitHub Integration: Actions, Issues, PRs, Discussions
  • Safety First: Sandboxed execution, minimal permissions, safe output processing
  • Cost Controls: Per-run AI credit budgets, OpenTelemetry cost analysis

Security Architecture (Layered)

  1. Read-only token — agent cannot push commits or write issues directly
  2. No secrets in agent runtime — credentials in isolated downstream jobs
  3. Sandbox + Agent Workflow Firewall — container with network restrictions
  4. Safe outputs gate — validates actions against policy before application
  5. Threat detection — scans proposed outputs
  6. Compile-time validation — schema validation, expression allowlisting, action pinning

gh aw CLI

The gh aw CLI compiles markdown to hardened .lock.yml GitHub Actions workflows running AI coding agents in containerized environments on schedule or manually.

Continuous AI

Agentic Workflows augment deterministic CI/CD with “Continuous AI” — event-triggered and scheduled agent jobs for triage, docs, testing, and cross-repo coordination.

Developed by GitHub and Microsoft Research (GitHub Next).