The agent harness that treats its own code as mutable state
Exo gives AI agents read-write access to their own source code, prompts, and runtime so they can evolve their own architecture at runtime.

What it does
Exo is an agent harness that hands the model the keys to its own house. It mounts its source code inside a sandbox where the agent can rewrite prompts, tools, adapters, and even core policy, then rebuild and restart itself. An immutable external log of events and conversations acts as a ground truth, surviving sandbox rewinds so the agent doesn’t erase its own history while experimenting.
The interesting bit
Most agents improve by updating memory or adding skills; Exo lets the model refactor the harness itself. The authors pitch this as a minimal, Bitter Lesson aligned bet: instead of hardcoding ever more elaborate scaffolding, give future smarter models full runtime control of the architecture and let them iterate on system-level properties like cost or scheduling strategy.
Key highlights
- Full recursive scope: the model can rewrite prompts, memory, tools, adapters, and core harness policy, then rebuild and restart.
- Sandbox with rewind: experiments happen inside a container that can be snapshotted and rolled back without touching the canonical event log.
- Multi-channel by default: supports CLI, web chat, and adapters for Discord, IRC, Slack, WhatsApp, and Signal.
- Lineage management: agents can clone themselves and maintain a family tree of variants.
- Canonical state lives outside the sandbox, so the agent retains a durable, append-only history across restarts and self-surgery.
Caveats
- Generalized GUI computer use is explicitly noted as missing, though the authors suggest asking the agent to build that capability itself.
- The project is still early stage; autonomous self-maintenance, failure recovery, and multi-agent orchestration policies are listed as active work, not finished features.
Verdict
Worth a look if you want an agent that can genuinely refactor its own scaffolding, or if you’re experimenting with long-running autonomous systems that need to cost-optimize or re-architect themselves. Skip it if you need a polished, batteries-included consumer agent today.
Frequently asked
- What is exoharness/exo?
- Exo gives AI agents read-write access to their own source code, prompts, and runtime so they can evolve their own architecture at runtime.
- Is exo open source?
- Yes — exoharness/exo is open source, released under the MIT license.
- What language is exo written in?
- exoharness/exo is primarily written in Rust.
- How popular is exo?
- exoharness/exo has 1.4k stars on GitHub and is currently cooling off.
- Where can I find exo?
- exoharness/exo is on GitHub at https://github.com/exoharness/exo.