An open-source coding agent rewrites its own harness and keeps a signed rollback record

INIT Labs describes the harness as the prompts, tools and loop around a model, and says INIT rewrites its own, tests each change and swaps it in mid-conversation. A component it calls the Root holds the API keys, the signed record and the undo, and the model cannot write to it. The page says INIT v0.3.3 has 136 files, of which 7 are prompts and skills, 5 are tools and 124 are compiled core, and that three changes have been made to them: two kept (October 20 and October 24) and one rolled back (October 21).

The vendor reports these figures, with the logs linked from its page: a core swap in 4.3 seconds, measured end to end with a simulated model; 87.9% of the prompt cache kept across a swap, from a live run on DeepSeek; and a first core change in 13.1 minutes with all checks included. A candidate change must pass four gates, with build and tests taking 439 seconds in the example shown. For the one rollback on record, the page says a change to the shell tool scored 5 points lower, inside the noise, and the gate refused it. The change was rolled back 9 seconds later, and the signed record shows sequence numbers 88 to 90. The page itself notes this was "refused, not proven harmful."

Only the project page has been read. The linked logs were not opened, and the page labels its 21-day comparison with a fixed harness "illustrative, not measured."

Why it matters: Most harness work today is a person editing prompts and tool definitions by hand. INIT gives a concrete design for letting the agent do that edit itself: a test gate, a twelve-signal list for rollback, and a record the model cannot alter. Whether the changes it keeps are improvements is the open question. The one refusal on record rested on a 5-point score difference the page itself calls inside the noise.