Turn engineering goals into measured pull requests.

Loop gives an agent a target, a working repository environment, and an acceptance gate. It changes the code and executes the result instead of stopping after a plausible patch.

Build failures route back into debugging. Test failures route back into implementation. Benchmarks decide whether a candidate is retained or reverted. The run continues without an engineer driving the terminal.

Your team receives the successful diff, the evidence behind it, and a pull request that enters the review process you already use.

One run, every decision visible.

Goal, current candidate, correctness gates, benchmark movement, and the resulting pull request in one operating view.

Acme Media / video-pipeline / Loop #1842 Example workspace

Reduce p95 decode latency below 40 ms

main @ 8d3e91a · started 09:14 · acceptance gate: identical decoded frames

Running · 07:42
Current p9536.8 ms−59.7% from baseline
Speedup2.48×target passed
Candidates124 retained · 8 reverted
Tests418/418correctness gate passed
Candidate historylower p95 is better
Iter.ChangeCorrectnessp95Decision
#12vectorize chroma copy418 passed36.8 msretain
#11parallelize frame queue3 failed34.1 msrevert
#10reuse decode buffers418 passed41.7 msretain
#09increase queue depth418 passed45.9 msno gain
#08remove redundant copy418 passed44.6 msretain
Execution pipelineiteration #12
Build
release + AVX2
00:41
Debug
asan · clean
01:08
Test
418 passed
02:36
Benchmark
30-run sample
03:17
Prepare PR
attach run evidence
now
Run activitylive trace
09:21:42Acceptance gate passed. p95 36.8 ms across 30 benchmark runs.
09:20:05Correctness confirmed. 418 tests passed; decoded frame hashes unchanged.
09:17:51Candidate retained. Replaced per-frame copy with vectorized batch operation.

Give your next task a complete execution loop.

Start with one repository and one measurable acceptance gate.

Connect a repository