Who verifies your AI verifier?

Your AI coding agent wrote 1,000 lines of code overnight. In the morning, you look at the commit history—all new features. Are you planning to verify it all manually? Then who's actually doing the work here, you or the AI?

Reticle wants to solve this. It's an AI verification layer built specifically for AI coding agents: it lets the agent check its own work before delivery. Founder Dibyanshu's motivation came from an incident—he gave an agent a prompt at night and woke up to find every feature broken. He tried every existing tool, found nothing that worked, and wrote his own.

How it works isn't complicated. Reticle instruments the code and watches network requests, application state, and console output. Each time an agent finishes a feature, Reticle pauses it, runs a verification pass, and hands the evidence back to the agent, which decides whether to ship or keep iterating.

The numbers from the official site:

• Of 88 manually injected bugs, Reticle caught 85; Playwright caught 59.
• Rerunning a 4-flow check suite costs Reticle about 480 tokens, versus roughly 120K tokens to re-drive it with an LLM—about 250x cheaper.
• More than 600K cumulative npm downloads, about 120K monthly downloads, and 6–7K users.

Usage is straightforward, too. Drop the GitHub link into an agent and say “install Reticle,” and it installs itself. As long as the agent supports the MCP protocol, OpenAI, Anthropic, and Gemini can all be configured. To try it, the official site is http://reticle.sh .

The first question this raises is: who verifies the verifier? One commenter wrote: “Then you need another AI to verify this verifier, and all I have left is the right to press merge.” Another developer was sharper: when tests are written by the same agent, it often ends up proving itself innocent. Checks should be generated from the specification before the code is written.

Reticle's response is open source. The SDK is Apache-2.0, and the verification engine is open source too. Its benchmark is reproducible, so anyone can run it. And replay is deterministic, not dependent on model judgment, which is more solid than using another LLM to check.

Robert Scoble published an interview with Dibyanshu on X and used AI to generate a summary of the key points. The video is here:

Reticle may not be the final answer. But it has grasped the real problem: AI writes code faster than humans can verify it, and continuing to make people stare at screens is a dead end. Letting agents verify themselves—even imperfectly—is better than not at all. You still need to keep an eye on it occasionally, though.

In the long run, who verifies the verifier? Open source is the short-term answer, and so is deterministic logic. But that question about “the AI that verifies the AI that verifies the AI” will always be there.