exoharness/exo

What it solves

Exo addresses the limitation of static AI agents by providing a framework for recursive self-improvement. While typical agents can update memory or skills, Exo allows an agent to have full visibility and write-access to its own source code, runtime logs, and harness policies, enabling it to evolve its own architecture and capabilities at runtime.

How it works

Exo operates as a complete agent harness with several key components:

• Host-side Loop: Manages user messages, tool execution, and context building outside the sandbox.
• Ubuntu Sandbox: A secure environment where the agent can install packages, run commands, and experiment, with the ability to snapshot and rewind state.
• Source Code Access: The project's source code is mounted directly into the sandbox, allowing the agent to read and modify its own logic, then rebuild and restart itself.
• Canonical State: A durable event log and history stored outside the sandbox to prevent recursive loops and maintain a record of experiments across restarts.
• Tools & Adapters: Functions for interacting with the world (e.g., shell commands) and long-running processes for external communication channels like Discord, WhatsApp, and Slack.

Who it’s for

It is designed for developers who want to build long-lived agents that can autonomously optimize their own system-level properties, cost-optimize themselves, or solve complex problems by iterating on their own underlying framework.

Highlights

• Recursive Evolution: Can modify its own prompts, memory, tooling, and harness policy.
• Cloning: Ability to clone itself and manage a lineage of clones.
• Durable History: Uses a canonical event log to ensure stability during self-modification.
• Multi-channel Integration: Supports various chat platforms including IRC, Discord, WhatsApp, Signal, and Slack.
• Sandbox Safety: Employs a vanilla Ubuntu sandbox with snapshot/rewind capabilities.