INIT — the coding agent that rewrites its own harness

6 min read Original article ↗

An open-source coding agent from INIT Labs

It rewrites its own harness. Grows with every model.

The harness is everything around the model: its prompts, tools and loop. INIT rewrites its own, tests each change and swaps it in, mid-conversation. A Root it can never touch can undo any change.

code the model may rewritea change in flightthe Root: never rewritten, undoes any change

diagram gen 41 cells rewritten 0 rolled back 0 Root moved never

+01THE AGENT, TAKEN APART

Take an agent apart. One layer it can never touch.

An AI agent is a model plus a harness.

INIT may rewrite every layer. Except one.

A new core, through four gates. Same conversation.

If a change does harm, the Root rolls it back.

A fixed harness becomes a ceiling. INIT grows.

one cell = one real file · INIT v0.5.6

  1. Assembled: the model, a frosted plate with no files, sits on four harness plates (prompts and skills, tools, compiled core, kernel), all standing on the amber Root.
  2. Exploded: every layer lifts away from the Root. INIT may rewrite the harness; the kernel only at higher autonomy, and still checked. The Root never moves and is never written by the model.
  3. Rewrite: a candidate copy of the compiled core slides into its own lane. One real file, src/loop/mod.rs, lifts out and is rewritten, rides through four gates (build and tests 430 s, self-map 1.0 s, boots beside the live one 19 s, can still edit itself 331 s) and drops back into its slot: swapped in within 4.1 s, in the same conversation.
  4. Rollback: a change to src/tools/bash.rs turns red and is refused; the Root's snapshot of the old file rises back into its slot, and the signed record ticks seq 88, 89, 90 (2026-10-21).
  5. Grows: beside a fixed harness, whose growing model overhangs a hard ceiling, INIT's harness widens to fit each new model. The Root stays the same size, in the same place.

Diagram. One cell is one real file of INIT v0.5.6: 9 prompt and skill files, 5 tool files and 129 compiled-core files. The kernel is drawn as a texture, not counted. Gate and swap times are measured on v0.5.0 with a simulated model, and the rollback is a real signed record (2026-10-21, seq 88 to 90). Motion, and all of step 5, are illustrative.

+02THE BODY

This is INIT's actual body. One cell per file.

The Body · live data

This is INIT's actual body.

One cell per real file at v0.5.6, the current release. Coloured: files INIT changed itself.

143files

532ktokens

3changes

prompts & skillstoolscompiled corekeptrolled backin flighthover a cell

ROOTkeys · the record · undo · the off switch — never written by the model

+03THOUGHT EXPERIMENT

A fixed harness pays for each friction every day. INIT pays once.

thought experiment · illustrative, not measured

Same job, same models, three weeks. The fixed agent trips on the same problems every day. INIT fixes each one once.

useful workfriction, paid againin flight: changing itselfkeptrolled back

4frictions the fixed harness still pays, every day, on day 21

0for INIT: it paid each one once

day 21 surprises nobody planned for 2

+04EVERY CHANGE CAN BE UNDONE

Freedom, with an undo button.

The model may rewrite almost all of itself. One thing it can never touch: the Root, which holds the keys, the record and the undo.

Where the line is · the model may rewrite

prompts & skills · tools · compiled core · kernel

sandboxed · no API key · one socket to the Root

every model call, every change ↓ metered and recorded

ROOT · nobody rewrites

API keys · signed record · snapshots + undo · harm signals · stop · resource floors

separate service · outside the sandbox · holds the signing key

+05THE NUMBERS

Seeing is believing. Each number links to its log.

The two DeepSeek numbers are live runs on v0.1 and v0.3; the rest were measured on v0.5.0.

87.9%

of the prompt cache kept when it swaps its core, so the chat carries on

40s

from install to first reply on a fast link, Rust toolchain included

$0.0005/turn

per live DeepSeek turn

4.1s

to swap in a new core, mid-conversation

13.1min

for its first core change, every check included

2,022tests

passing in the release's CI

Real model · gated mode · 2026-10-21

A rollback, signed.

  1. 1A change to its own shell tool scored 5 points lower, inside the noise.
  2. 2The gate refused it.
  3. 39 seconds later it was rolled back. Every record is signed.

Refused, not proven harmful · why refuse it? →

The change scored lower.

seq 88 · sealed_g4by root · 08:32:33.794Z

candidate2efd6c39…d_permille−50statusinferior

prev ← seq 87sha256 chained25519

+0.063 s · hash links to seq 88

The gate refused it.

seq 89 · gateby kernel · 08:32:33.857Z

verdictrejectreasonscore_inferior

prev ← seq 88sha256 chained25519

+8.722 s · hash links to seq 89

It was rolled back, and signed.

seq 90 · body_rollbackby kernel · 08:32:42.579Z

fromc88be5f…toaaf07b7…

prev ← seq 89audit-verify: lineage oked25519

lineage.jsonl records 88–90 · fields verbatim, abridged · m7-results.md

+06RESEARCH

Four open problems. One instrument to study them.

INIT Labs builds the agent and studies what it does to itself. Every change it makes is signed, tested and undoable: a dataset for the questions below.

nestfoodtrail, laid by antsant carrying food

R1 · Swarms · open

Colonies, not committees.

Today's agent "swarms" are a manager with assistants. A bee swarm has no manager.

Open: can agents following only local rules solve what one agent, or the best of many, cannot?

genesmutationcrossoverdied out

R2 · Evolution · open

Agents with DNA.

From agents that edit their own code to agents that inherit, recombine and evolve, with no human-written score.

Open: is there a genome for agents, where mutation makes variation instead of collapse?

the agentits forecast of itselfa failure it saw comingone it missed

R3 · Metacognition · open

An agent that sees its own mistakes coming.

Today's agents notice an error after they make it. Minds forecast their own failures, and change course first.

Open: can an agent model itself well enough to foresee its own failures, and act on that in time?

new changedetectorslipped through

R4 · Immunity · open

An immune system for self-improving AI.

An agent that upgrades itself can be infected by its own upgrades. Nothing guards against that yet.

Open: can it tell self from non-self in its own changes, without attacking the healthy ones?

Why INIT Labs

The record is the dataset.

Every change INIT makes to itself is signed, tested and undoable. Each one is a data point for all four questions.

signed · gated · reversible

Open problems, not results. We will publish what fails, too.

+07TRY IT

One line to install.

$curl -fsSL https://github.com/nagisanzenin/INIT-Agent/releases/latest/download/get.sh | sh

then, in a new terminalinit-agent --workspacekeeps it to one folder ·read get.sh first ↗