Silex-Research/DontPanic

DontPanic

Trust infrastructure for AI coding agents.

AI agents can write real software now. The open question isn't whether they can produce code — it's whether you should ship what they produce without looking.

DontPanic is the layer that makes that decision safe. It pins an agent's work to a plan it can't quietly rewrite, has a different AI model review the result, stops and asks a human before anything risky lands, and keeps the receipts.

Think of it as code review, CI, and an approval queue — but built for the case where the author is an AI.

Never let the same AI that wrote the code be the only AI that approves it.

It works with Claude Code, Codex, Gemini, Grok, local models, OpenClaw / Hermes workflows, and any MCP-enabled tool.

Status: public alpha. Ready for source installs by people comfortable with local agent CLIs and command-line work.

The problem

An AI coding agent is fast, confident, and tireless. It is also happy to:

• ship broken code and tell you the tests passed
• review its own work and approve it
• miss the security assumption nobody wrote down (the webhook nobody verified)
• spend $40 looping on a test that never goes green
• leave nothing behind that explains who decided what

For a prototype, none of that matters. For software other people depend on, all of it does. DontPanic lives in the gap between two sentences:

"The agent says it's done."

"Should we actually merge this?"

Why this gets harder as the work gets real

Vibe coding is great for demos and side projects. Production is a different animal: hidden dependencies, security requirements, deploy steps, cost limits, and a person who needs to know who approved what.

A strong engineer still gets peer review, QA, a budget, and a deploy gate — not because they're bad at the job, but because complex systems need more than one set of eyes. Autonomous agents are the same, only more so.