Your agent is about to refactor this. Do you know what it does right now?
Baseline drives your running app and records what it actually does. The agent
refactors. Baseline re-drives it and tells you exactly what changed.
By execution. It never reads the diff.
The problem isn't that agents write bad code
It's that they change behaviour you didn't know you depended on, and your test suite
says nothing — because the tests were written against the code, not against what the
software actually does in front of a user.
A study of 7,191 agent-authored pull requests found agents are safer than humans
on new features (2.89% vs 7.74% breaking changes) and worse on refactoring
(6.72% vs 4.36%). Refactoring is exactly the task where every check stays green.
“I let a popular agentic ai refactor some code for me today, it looked
really nice, but despite the fact that it was meant to be splitting a file full of functions
that already compiled and worked it rewrote them all and broke them comprehensively.”
“It would try to make a change, run tests, realize it broke something
somewhere else, try to fix that, cause another issue… decide to make the tests pass by
removing the tests!”
“It did the job very well. It actually did it too well, except for one
small footgun… a day passes… I bisected the game code with git, went back two months, and
everything was broken.”
“I spend a week troubleshooting a subtle resource leak on a commercial
project that was introduced during a vibe-coding session… The bug caused an all-stop and a
week of lost effort.”
These are public comments about the problem, quoted and linked with
attribution. They are not endorsements — nobody above has seen this page.
How it works
Record Point it at your app running on localhost. It explores, drives the real
interface, and records what actually happened — responses, state, rendered output, side effects.
Let the agent work Refactor, restructure, port, upgrade a dependency. Baseline
doesn't watch, doesn't read the diff, and doesn't care what changed in the source.
Re-drive and diff Same inputs, same paths. You get a list of behaviour that
changed — with the exact input that produced it. Not a pass/fail on someone's opinion of
your code.
What it deliberately is not
Not a test generator. There's no suite to maintain. The artifact is a
verdict on one change, and then it's gone.
Not a code reviewer. It never sees your implementation — so it can't be
fooled the way “ask the model to write tests for the code the model just wrote” is fooled.
Not a SaaS you send your code to. Local CLI plus an MCP server. Nothing
leaves your machine, so there's nothing for anyone's security team to review.
Not a staging-environment product. No script tag to install, no eBPF, no
containerising into someone's simulator, no production traffic required. Your app on
localhost, right now, is enough.
Would you use this?
I'm building this only if the pain is real. If it is, tell me — and tell me the part I've
got wrong. I read every one of these.