Agent Architect Kit

Agent Architect Kit

**Version 1.4.0** · See [CHANGELOG.md](CHANGELOG.md) for what's new.

**New here? Start with the [Agent Starter Kit Pro](https://ultrathink.art/items/250)** — the canonical free starter kit. This repository is its extended superset: six role definitions, deeper process docs, an auditable task queue, a dual-harness worker, health monitoring, and launchd automation.

A free kit for running structured Claude Code or Codex CLI agents on one codebase. It is extracted from a production system that has processed 5,000+ autonomous tasks.

The problem

Unstructured agents fail predictably:

• Sessions forget prior mistakes.
• Agents self-report success without reproducible evidence.
• Parallel work collides.
• Safety and deployment rules drift between sessions.
• A client default promises QA, but another client or raw API silently skips it.
• A README says a harness is portable while the launcher cannot actually run it.

What you get

• First-class Claude Code and Codex scaffolds — each uses the project and custom-agent conventions its harness actually reads.
• Six roles — coder, QA, designer, product, marketing, and operations.
• One dual-harness worker — select claude or codex through AGENT_HARNESS; Claude remains the backward-compatible default.
• Protected prompt delivery — task and brief text goes through stdin via a mode-0600 temporary file, never shell interpolation or command arguments.
• Fail-closed preflight — executable, project instructions, role config, sandbox, and permission settings are validated before the orchestrator claims a task.
• Auditable QA handoffs — coder/product opt-outs require a persisted reason, and an executable server reference enforces an idempotent QA child again at completion.
• Integrity verification — checksums.txt and full/scripts/verify-kit prove whether a source tree or download was modified.

agent-architect-kit/
├── core/
│   ├── CLAUDE.md