Init – a Rust agent harness that recompiles and hot-swaps its own core mid-task

INIT is an open-source Rust coding agent built around the idea that an agent’s harness - its prompts, tools and code - should adapt along with its model. During a conversation, INIT can rewrite parts of that harness, compile a candidate core and swap it into the running agent. Changes pass build and test checks, a self-map check and a trial boot beside the live version. A separate, unmodifiable Root holds credentials, signed logs, snapshots, rollback controls and stop mechanisms; it can reject or undo changes that trigger harm signals. The argument is that a fixed harness repeatedly incurs the same friction, while a self-improving one can fix problems once and accommodate more capable models without surrendering control.

Reported measurements include a 4.3-second core swap that preserves 87.9% of the prompt cache, a 13.1-minute first core change including checks, and 1,897 CI test runs with no failures. In a recorded example, a shell-tool change scored five points lower, was rejected as inferior and rolled back nine seconds later; signed, linked records document the sequence. These results demonstrate a working mechanism, not proof that self-modification is broadly safe or beneficial. INIT Labs presents the system as an instrument for investigating agent swarms, evolutionary coding, self-prediction of failures and defenses against harmful upgrades. Those remain open research questions, and the project says it will publish failures as well as successes.