first-fluke/oh-my-agent

What it solves

Many AI agents claim success based on their own internal narration, which can be misleading or contradictory. oh-my-agent solves this by implementing a mechanical verification system that checks actual artifacts (like test results and files) rather than trusting the agent's summary, ensuring that work is actually completed and verified.

How it works

The project provides a multi-agent harness that organizes specialized agents (e.g., backend, frontend, QA) into a virtual engineering team. It uses several mechanical gates to ensure quality:

• Stop-hook gates: Prevents a session from ending until specific scripts (like typecheck or test) exit successfully.
• Anti-Circumvention Gates: Verifies the existence of specific artifacts (like plan JSONs and QA result files) to ensure workflows were actually run.
• Independent Judges: Spawns separate agents with fresh contexts to re-verify criteria without being biased by the implementer's claims.
• Event-sourced state: Records every decision and gate outcome in an append-only JSON log for auditing.

It projects a single source of truth (.agents/ directory) across various agent runtimes (Claude Code, Cursor, Codex, etc.), allowing users to switch vendors via configuration.

Who it’s for

Developers and engineering teams who want to move beyond simple chat-based AI coding and implement a rigorous, multi-agent workflow with automated verification and cross-vendor flexibility.

Highlights

• Role-based Architecture: Specialized agents for architecture, DB design, frontend, backend, and security.
• Multi-Vendor Support: Works across multiple AI runtimes including Claude Code, Cursor, and Qwen Code.
• Mechanical Verification: Uses hard exits and file checks instead of LLM-based "looks correct" assessments.
• Token Efficiency: Implements a two-layer skill loading system to reduce context window usage.
• Monorepo Awareness: Automatically detects and routes agents to specific workspaces in pnpm, nx, turbo, or lerna projects.