A recursively self-improving AI agent harness for scalable, safer agent applications

Build an agent system whose executive can inspect and modify its own prompts, context assembly, tools, skills, adapters, and other policy code at runtime. Separate stateless policy from protected state and an isolated execution environment so the system can evolve while preserving history, protecting secrets, testing changes, and rolling back failures.

From Latent Space — Exo: Harnesses should see their own code and logs — Alex Krentsel at 01:55

Problem: Conventional agents can modify memory, skills, or predefined extension points, but cannot recursively improve the full machinery that constructs context and executes actions. They also commonly mix agent policy, secrets, state, and execution in one environment, making self-modification risky and large-scale deployment difficult.

For: Companies building agents for many users or ongoing usage streams, and developers who need agent systems that can improve task performance or reduce inference costs without manually redesigning the harness.

Examples

Exo playing Pokémon: Exo inspected the game's RAM, mapped information such as world position, active Pokémon, and battle state, modified its own game integration, and fed the resulting information into its system message.

Soon you can unlock the full business plan.

Behind this: 13 build steps · 3 tools and how each is used · how to validate demand · 3 more real examples · 8 things the video never answers.

Inquire for details

Other takes on AI agent platforms and compute access

All AI agent platforms and compute access ideas →