Introducing Vet: catch your coding agent’s mistakes Open source, fast, local. Zero telemetry Can review full PRs and it tells us when Claude/Codex/OpenCode lies about passing tests it never ran 👿 We use it daily and it saves tons of headache

We built Vet because our coding agents would constantly implement a feature, hit a wall, and quietly stub things out with hardcoded data. The code looks fine and tests might pass, but it's not what we asked for. We run Vet in agent loops, drop it in CI, and invoke it from the CLI. Get the code:

Vet catches logic errors, unhandled edge cases, silent failures, insecure code, scope drift, etc At Imbue, we believe open agents must win over closed platforms for humans to live freely in our AI future. Vet is one of many upcoming projects toward that end. Read more:

Woohoo, proud of Andrew and the team for building and shipping this! It's sooooo useful

Caught my agents hardcoding expected outputs just to get green checks, more times than I'd like to admit. Code review doesn't always catch it because the logic looks right. What's review latency on a full PR? If fast enough to run mid-agent-loop that changes the workflow completely.

Review latency on a PR is generally less than a minute, but running Vet in the loop, as the agent works, will be able to catch hardcoded credential issues before any code can be seen by other people! We run Vet in the loop because of issues like this so it can ensure agent changes match requests.

Had so much fun making this one and highlighting the work Andrew and the team put into this product!

vet is carrying my entire workflow rn and half of you don't even know it exists yet. That's fine. More alpha for me. Seriously though if you're "in AI" and haven't touched vet you're cosplaying. Sorry not sorry.

So excited about this! And really cool graphics!!

I've been using Vet as a final pass code review, and despite a solid pipeline of linting, tests, ratchets and agent directives it still regular catches issues with agent generated code. It's indispensable.

Vet it and forget it — don’t wait till it’s too late! 🤷🏻‍♂️

It's evolution-powered code review in your agent loop!

so hyped!

Great work all around 👏

This video rules @nikhilville

gaslight detection in an easy to install and use form factor!

@nikhilville is COOKING

Verify everything feels like the new rule in AI workflows

Game changer for coders, ensuring accuracy 🚀

This is the missing layer. Most frameworks let the LLM claim "tests pass" without independent verification. Vet moves the check from probabilistic to deterministic — the model says it works, the tool verifies independently. That boundary is what matters.

I keep a regression log beside the code for this reason. Tests catch known failures. The log preserves the weird ways an agent previously declared victory while missing the actual request.