Platform Agent Harness

platform-agent-harness

An issue-driven, agent-first development harness for platform engineering teams running one or more Claude Code agents on infrastructure repositories — without making chat history, one machine, or one agent's memory the source of truth.

Adapted from an agentic development coordination system originally built for game engine development. This port applies the same coordination principles to platform engineering work: Kubernetes, Terraform, observability pipelines, secrets management, CI/CD, and internal developer tooling.

---

What this solves

Platform infrastructure repositories have a specific challenge that product feature repos don't: **changes here affect every engineering team in the company.** A misconfigured Terraform module or OIDC trust policy doesn't break one feature — it can break all of production.

Running AI coding agents on infrastructure without coordination is dangerous. This harness provides:

• Issue-driven coordination — every task is a GitHub issue with type, priority, lane, and blast radius before any agent touches it
• Coordinator/worker role separation — one agent manages the backlog, reviews PRs, and owns the shared foundation; workers implement within clearly bounded lanes
• Atomic claim protocol — agents can't collide on the same work; the claim is visible and timestamped on the issue itself
• Shared foundation protection — OIDC trust, Terraform module interfaces, base images, and SLO frameworks are coordinator-gated with no exceptions
• Competing-model review — PRs get an independent security and reliability review from a different AI provider before merge
• Skills library — agents load the right guidance for their specific task before writing anything

**GitHub is the coordination layer.** All decisions, claims, and handoffs are on issues and PRs — not in invisible conversation. Any agent on any machine can read the state and continue safely.

---

Repository structure