CircleCI
The continuous integration platform for the World’s best engineering teams
Start new thread
Olaf Molenveld

Chunk sidecars - Validate agent-generated code before it ever reaches CI

by
CircleCI
•
AI agents write code fast. Validation still happens after the push — by then the context is gone. Chunk sidecars run scoped microbuilds before commit, in a real CI mirror. Auto-detects your stack. ~27s average vs ~5 min billable compute for a full run. 3x–5x fewer tokens in retry loops. If something fails, the agent iterates before anything reaches shared CI. Run chunk init. Works with Claude Code, Codex, Cursor, or custom agents. Free for all CircleCI users.

Add a comment

Replies

Best
Olaf Molenveld
Maker
📌
Hey Product Hunt 👋 One thing that became obvious while experimenting with agent workflows internally: the code generation is becoming the easy part. The expensive part is when validation happens after the push, CI fails, and now the agent has already moved on and lost the useful context around the change. We’re also seeing this show up in the broader data. In our 2026 State of Software Delivery report, build volume increased 59% year over year while main branch success rates dropped to a 5-year low. That’s why we built Chunk sidecars. They run lightweight microbuilds against environments that mirror your CI stack while the agent is still iterating on the code. Feedback comes back in under 60 seconds, so failures get caught before they ever hit shared CI. Here’s a short demo: https://circle.ci/4dq9fph Would love feedback from people already experimenting with AI-assisted development workflows!
GuoleiZuo

This is a strong wedge if the sidecar output is easy to review inside a normal PR. For agent-generated code, the hard part is not only catching failures, but making the failure cheap enough that developers actually read it before CI.

I’d be curious whether Chunk sidecars can show “why this changed” next to tests, touched files, and risk areas, rather than only a pass/fail validation result.

Olaf Molenveld
Maker

@studentzuo Tnx for your great thoughts and feedback! The current premise for Chunk sidecars is that it can validate and trigger for (agent)fixes autonomously and transparently, before even hitting the CI, and thus not polluting the PR with unnecessary info for a human reviewer. Though I can imagine that passing some kind of summary/event-log to the "outer loop" CI pipeline with issues and fixes might be valuable for a PR review. Thanks for sharing!

Aiswarya Subramanian

Very cool, super important given all the issues we run into after pushing agent written code.

Congratulations @olafmolenveld @z00b , upvoted :)

So how exactly do you test before deploying? Do you show which lines are risky or something, if so - how do you figure that out?

Olaf Molenveld
Maker

@aiswarya_s Great question, thanks for the upvote! :)

Chunk sidecars doesn't identify risky lines or do static analysis. What it runs is what we call a microbuild — a scoped version of your CI pipeline in a cloud environment that mirrors your actual CI stack. "chunk init" auto-detects your stack and configures what to run. That's not just tests — it's linters, build steps, and whatever checks are relevant for the change, running against the same environment your full CI would use.

The difference from just running tests locally: your laptop doesn't catch environment-specific failures, missing service dependencies, or build issues that only surface in CI. The sidecar does, because it's an actual CI mirror.
The idea is that catching a failure 27 seconds into a microbuild costs a lot less than catching it 5 minutes into a full pipeline run — and more importantly, the agent still has the context to fix it.

Does that answer what you were asking? Happy to go deeper on how the environment detection works if useful.