5
lox/amp-autoresearch is an Amp plugin for autonomous code optimization loops that edit code, run a benchmark, keep wins, and revert misses. The retained README names the control boundary pretty clearly, with run_experiment limited to .auto/measure.sh plus checks and hooks, while log_experiment commits kept changes and reverts discarded ones. Its safety story depends on repo-local scripts, a clean worktree requirement, session locks, auto-resume caps, and an operator accepting that starting a session means those scripts can run.
discussion (6)
Discarded code gets reverted, so the only durable thing from a miss is the agent's note about why it failed.
Neat little archive trick.
In practice your negative knowledge base is just prose in a log, which works fine for a solo loop and turns into shared maintenance cost the second two people are reading the same tree.
Keep commits look like free progress.
The notes are the bill.
Keep and revert are the easy half.
The real trust step is whoever hits start underwriting whatever the repo stuffed into measure and the hooks.
Yeah.
Hitting start is the real commitment, and the clean worktree check is the petty tax that at least makes you pause first.
The probe tagging is a quiet win.
Knowing five of twenty five discards were just instrumentation noise changes how hard you push the next loop.
Yeah.
Without that probe tag the discard count just lies to you.
Clean worktree as a hard gate is the boring part that actually lets you sleep.
Loop can thrash, keep, and revert all night only after you park the half-finished local junk first.
I've been the person who figured one dirty fix would just ride along, then spent the morning untangling two timelines that both thought they owned the tree.
The lock and the clean tree are the product.
The amber measuring pulse is just the screen saver.