A Berkeley paper has agents rewrite their own harness during the task and reports lower model cost

The authors describe Exo, a runtime in which the agent revises its own harness policy (context construction, tools, memory) while it works, using the same code-editing tools it uses for ordinary tasks. A small protected layer keeps execution history and controls when a new policy goes live. The authors report that on long-running objectives in an emulated RuneScape environment, Exo cut model cost by 12 to 53 percent against Codex and Pi in standalone runs and by 45 to 82 percent in transfer runs, "while matching task completion rates." Against the same starting policy without the ability to edit itself, they report 35 to 61 percent lower held-out cost on AppWorld and SREGym with comparable correctness. On Continual Learning Bench coding tasks, cost including learning fell 14 percent, with both policies resolving every issue.

These are the authors' own results on a game environment and two benchmarks, with models the paper names as gpt-6.1-sol and gpt-5.6-sol. The coding result is the smallest of the three. We read the evaluation sections; we did not find a code release or an independent replication.

Why it matters: Most harness tuning today is a human editing prompts and tool lists between runs. This paper moves that edit inside the run and keeps a protected layer that decides what activates. If the result holds, the audit trail of an agent's self-edits becomes something a team would need to review. The paper's transfer runs, where the agent reuses a harness it already specialized, account for the largest savings it reports.