AgentForge is a control system for AI coding agents. Instead of hoping your agent "writes good code," you give it structured engineering workflows — the same ones senior engineers actually follow.
What it is:
28 skills covering the full lifecycle: define, plan, build, verify, review, ship
Each skill is a structured process with steps, anti-rationalization tables, red flags, and verification gates
Works with Claude Code, OpenCode, Gemini CLI, Copilot, Cursor, Windsurf
Why it exists: I've watched agents confidently ship broken auth, skip error handling, and deploy on Friday afternoon because nobody told them not to. These aren't prompt engineering tricks — they're encoded workflows. The test-driven-development skill makes the agent write a failing test before touching implementation. The shipping-and-launch skill forces a rollback plan. The doubt-driven-development skill makes the agent challenge its own assumptions before continuing.
What's different:
Anti-rationalization tables: Every skill lists the excuses engineers use to skip best practices ("CI is too slow" → "Optimize the pipeline, don't skip it")
Verification gates: Checklists the agent must complete before proceeding
Cross-skill consistency: A quality gate ensures all 28 skills follow the same anatomy and reference each other correctly
The catch: This isn't magic. The agent still needs to follow the skill. But when it does, the output is consistently better — fewer "it works on my machine" patches, more actually-shippable code.
Repo: [https://github.com/borhen68/SkillEngine]
I'd love feedback from anyone running AI agents in production. What's the most expensive mistake your agent has made?
Top comments (0)