AutoProteus.

AN EXPERIMENT IN AGENT EVOLUTION

A small seed.
Room to grow.

An AI agent that can edit its own tools.

AutoProteus starts with a small, model-agnostic harness. Give it a task, explore the changes it makes, and test what comes next.

Explore the project See how it works

Experimental project · Local or hosted models · Source on GitHub

FIELD NOTES / 001
A seed with room to grow A central seed connects to three branches representing tools, tests, and a new candidate. A surrounding boundary keeps the experiment contained. NEW TOOLTESTSCANDIDATEREVIEWSEED
Small by design.
Editable by the agent.AP / SEED
A small, readable seed Your choice of model Changes you can inspect

THE APPROACH

Let the agent change.
Keep the controls outside.

The interesting question is what an agent can build from a simple starting point. AutoProteus separates that editable seed from the controller that runs the experiment.

ANATOMY OF A RUNVM MODE
HOST CONTROLLER

The experiment’s controls

Model connection · Interface · Execution policy · Recovery

Model responses & saved artifacts
DISPOSABLE LINUX VM

The editable seed

Agent loopFile toolsSession stateSeed tests

The agent works on its own copy of the harness.

In VM mode, host drives are not mounted and the guest has no network adapter. See the project’s security documentation for the full boundary and its limits.

FROM TASK TO CANDIDATE

A change is a starting point.

Inspect it. Check it.
Then decide what to try.

  1. 01

    Give it a task

    Connect a local model or a compatible hosted provider. Start with ordinary file work, or ask the agent to improve a concrete part of its seed.

  2. 02

    Inspect the edits

    Read the source changes, saved files and session history. The agent’s copy can evolve while the protected starter remains available.

  3. 03

    Check the result

    Compare the edited seed with the starter in fresh VMs. Existing task checks help catch regressions; new features need their own review.

  4. 04

    Try the candidate

    Choose to try the edited seed in a new VM, or return to the starter. A model-written change is never activated automatically.

A WORK IN PROGRESS

Built for exploration.

AutoProteus is an experimental project. The goal is to study useful, inspectable changes to a small agent harness. Results depend on the model, task and configuration.

Read the project notes
Can I use a local model?

Yes. AutoProteus supports local models through Ollama and compatible hosted chat services. The model connection stays outside the editable seed.

Where do I start?

The documented VM setup runs on 64-bit Windows 11 with Windows Hypervisor Platform enabled. The launcher guides you through preparing the environment. See the README for requirements and other execution modes.

Can it reliably improve itself?

That’s the question being explored. A working candidate in a particular test does not establish reliable self-improvement. Source review, independent checks and repeat trials are part of the process.

FOLLOW THE EXPERIMENT

See what grows from here.

Explore the code, read the findings, and try the seed.

Open AutoProteus on GitHub