Exo

Exo is an agent harness architecture for runtime self-improvement: it collapses the optimization loop into task execution so an agent can inspect and edit its own harness policy while working. Paper: https://arxiv.org/abs/2611.00463 (arXiv 2611.00463); code: https://github.com/exoharness/exo . Authors include researchers affiliated with UC Berkeley, Andreessen Horowitz, and Braintrust.

Design (exokernel-inspired) splits the system into:

• An evolvable policy layer (executor): tool selection, context assembly/compaction, memory/skill retrieval, and when/how to improve further—all editable code;
• A protected mechanism layer: append-only execution history, versioned artifacts, credentials and sandbox APIs, plus an independent guardian that gates policy activation and enables rollbacks.

Reported results (paper): on long open-world game objectives, model cost drops about 12–53% vs Codex/Pi in standalone runs and 45–82% in transfer runs; vs the same starting policy without policy editing, held-out cost falls on AppWorld/SREGym with comparable correctness. Model weights stay fixed; improvements are harness code changes.

Problem

Harness policies are usually hand-written or tuned in an external loop that repeatedly runs candidates on benchmarks. New domains often lack benchmarks; long tasks make evaluation expensive; and fixed policies cannot absorb runtime lessons. Exo lets the running agent revise both task behavior and its improvement strategy with the same code-editing tools it already uses.

Safety mechanisms

• Storage contract: append-only history so self-edits cannot erase evidence;
• Replacement contract: candidates activate only if the guardian accepts them; independent rollback;
• Adaptations appear as inspectable code diffs.

Limits

• Fate-sharing risk if a bad edit breaks the agent—hence the protected core;
• Specialization has upfront cost and may not pay off on one-shot, low-reuse work (paper example);
• Numbers are from the paper's setups/models—validate in your environment.

Who it's for

• Teams researching or shipping self-improving agent harnesses;
• Long-horizon workloads where reusable tools and context policies amortize learning.

Getting started

1. Read https://arxiv.org/abs/2611.00463 for the policy–mechanism split and guardian;
2. Clone https://github.com/exoharness/exo and follow its README;
3. Keep mechanisms minimal and protected; mount editable policy in the sandbox; gate activation via the guardian;
4. Measure cost/history impact of self-edits; restrict self-editing on low-reuse tasks.