When your agent loops, unrut gets it out.

Coding agents get stuck repeating the same failed fix, burning your time and credits. unrut notices the loop, steps back, and restarts the work from a different angle.

We'll email you once, when the beta opens. No spam, no sharing.

Fig. 1: the rut, and the way outStatus: building the benchmark

Same fix. Same error. Again.

Once an agent commits to a wrong idea early, it tends to keep defending it. Each retry looks a little different, but it's the same approach, so it fails the same way. You either babysit it or come back to a long, expensive session that got nowhere.

attempt 4edit src/slug.ts ✕ expected "creme-brulee"

attempt 5edit src/slug.ts ✕ expected "creme-brulee"

attempt 6edit src/slug.ts ✕ expected "creme-brulee"

unrutsame failure 3×, stopping this approach

unrutrestart from clean files; ruled out: regex strip

attempt 7edit src/slug.ts ✓ all tests pass

Illustration of the intended behavior, not a recorded run.

Notice. Step back. Try another way.

Notice the loop

unrut watches the agent's edits and test results. When the same failure, or the same files, come back three times, it calls it a loop.

Step back

It pauses the agent and writes down the assumptions behind the failed attempts, so the next try can't quietly repeat them.

Restart clean

It resets to fresh files and starts a new attempt with what didn't work ruled out, instead of piling more patches on a bad idea.

We're testing it before we sell it.

Recovery is only worth paying for if it works. So we're building a benchmark of stuck loops: tasks where the obvious fix fails. Every recovery strategy is scored by one thing, whether the tests pass, and compared against doing nothing. We'll share the results, including the ones that don't flatter us.

HarnessBuilt. Runs agents, detects loops, scores each run by its tests. Each case ships with a plausible wrong fix that must still fail, so a lucky guess can't pass.
Test cases4 seed cases. Aiming for 15–20 drawn from real stuck sessions. slug-diacriticsutc-date-shiftconfig-merge-shallowmemo-cache-key
ResultsNot published yet. No numbers until there are real runs behind them.

Fair questions.

Which agents will it work with?: We're starting with coding agents you run yourself in a terminal or editor, where outside tools can watch and steer a session. Tell us what you use when you join, and we'll prioritize accordingly.
Won't recovery cost extra model calls?: Yes, a restart uses more calls. The number that matters is cost per finished task. A loop that runs for an hour is the expensive case, and that's what we measure.
Doesn't my agent already handle this?: Agents retry, but they usually retry the same idea. If your agent rarely gets stuck, you probably don't need unrut, and the benchmark will tell us how often that's true.
What will it cost?: The beta is free. We'll set a price only after we can show what it saves.

Stop babysitting your agent.

Join the waitlist and we'll email you when the beta opens.

We store your email only to tell you about unrut, never share it, and delete it on request: gavin@unrut.dev.