Exo Lets Agents Edit Their Own Harness Mid-Task

Training an agent to get better usually means stopping the work, gathering a representative dataset, and burning through expensive runs before anything improves. Exo moves that loop inside the task itself, with the agent editing its own harness code while it works. It is a research architecture from a paper, not something you install today.

On AppWorld and SREGym, that self-editing cut held-out model cost by 35 to 61% against the same starting policy without it, and correctness stayed about the same. Some tasks also finished faster.

What Happened

Exo is an agent runtime architecture for in-harness learning and adaptation during ongoing work, presented in a paper on building self-improving harnesses.

Normally, harness optimization runs offline. That means you need a representative dataset and many expensive runs before the agent improves at all. Exo runs that same loop during normal task execution instead.

The agent edits its own harness code as it works. A small protected layer sits around those edits: it keeps the execution history and controls which policy version is active, so every change can be inspected rather than quietly taking effect.

The measured results, against the same starting policy without self-editing:

• 35 to 61% lower held-out model cost on AppWorld and SREGym, with comparable correctness.
• 13% lower mean task time on AppWorld.
• 35% lower median diagnosis time on SREGym.

Why It Matters

Offline harness tuning is the slow, costly part of agent work: you pay for runs before you know whether the change helped. Moving the loop into the task means the improvement happens while the work is being done, and the protected layer keeps a record so you can still check what changed and when.

If you build or tune agents, this is worth reading for the architecture alone, especially the idea of a small trusted layer guarding policy versions while everything else is fair game for the agent to rewrite.