Exo lets AI agents rewrite their harness while working

AI agents depend on more than a language model. Their surrounding software, often called a harness, determines which tools they can use, how they assemble context, what they remember and how they execute actions. In the Exo preprint, Alexander Krentsel and colleagues investigate whether agents can improve this machinery during ordinary work rather than relying entirely on developers or a separate optimization process.

The distinction matters: Exo does not retrain the underlying model. Its adaptations appear as changes to code, prompts and reusable skills. The running agent can inspect its harness, edit it with familiar coding tools and request activation of the revised version. It can also modify the procedures that determine when and how it attempts further improvements.

An editable harness above a protected core

Allowing an agent to rewrite its own operating software creates an obvious problem: a faulty edit could break both its ability to work and its ability to repair itself. Changes could also destroy the records needed to understand what went wrong.

Exo addresses these risks by separating editable policy from protected mechanisms. The editable layer controls tools, context construction, memory retrieval and model calls. The protected layer maintains an append-only execution history, retains earlier versions of stored artifacts and governs activation of replacement policies through an independent component called a guardian.

This separation is intended to let the agent change its behavior without erasing its past. Execution records include model costs and token usage, giving it evidence for identifying expensive procedures. Developers can inspect the resulting changes alongside the history that motivated them.

However, these protections are narrower than a guarantee of safe or successful behavior. The default guardian checks whether a candidate builds successfully, not whether it performs correctly or improves efficiency. The preservation guarantees require storage and replacement controls to remain outside the agent’s editing authority; the paper explicitly says its broadly writable development configuration does not enforce those restrictions. Rollback also cannot undo external effects.

Where specialization paid off

The authors first tested Exo against Codex and Pi in an emulated RuneScape environment, using the same model and reasoning effort across harnesses. Agents pursued combat, crafting and currency objectives.

Exo completed the combat and crafting objectives at 12–53% lower model cost in standalone runs. When repeating those objectives on fresh accounts with retained experience, its model costs were 45–82% lower than the competing harnesses. All three completed the finite objectives. These open-world results came from one standalone-and-transfer pair per objective and harness, rather than a large repeated sample.

The combat run illustrates the proposed mechanism. Exo initially spent more while building reusable tools. It later revised those tools to skip stalled targets, retreat earlier and batch training actions. It also shortened the history included in context while retaining older strategy in durable memory. The initial investment eventually gave way to lower model spending.

For graded benchmarks, the comparison changed: Exo was tested against the same starting policy with policy editing disabled, rather than against Codex and Pi.

On AppWorld, both versions passed 96.9% of held-out tasks across three seeds. Exo reduced held-out model cost by 61% and mean task time by 13%. Including learning and reflection, total model cost fell by 43%.

On 20 held-out SREGym service incidents, median diagnosis time fell from 81 to 53 seconds, and mean model cost declined by 35%. The authors report comparable correctness across additional holdout runs. On a fixed sequence of 19 coding issues, repeated five times per arm, both policies resolved every issue; Exo used 14% less model cost including reflection. That coding experiment had no separate held-out phase.

Improvement needs opportunities for reuse

Self-editing did not always repay its cost. In a quest-focused game experiment, where tasks involved largely distinct steps, Exo cost more per quest point than either competitor. Additional AppWorld experiments with a smaller model produced limited editing and no reuse of its created tools on held-out tasks.

The preprint therefore supports a conditional conclusion: inspectable harness changes can reduce model spending when an agent discovers procedures worth reusing. It does not establish that self-modification reliably improves every workload, or that preserving history and checking builds is sufficient to ensure correct behavior.