Self-Evolving Agent Harness

The short version

The problem

Agents often need to adjust to new tasks after deployment, but updating the underlying model can be costly or impossible, and changes to prompts, tools, or workflows can accidentally make performance worse. The challenge is to learn from experience while carrying useful changes forward safely.

The idea

The paper treats an agent’s surrounding software—its instructions, memory, tools, and runtime—as something that can be revised while leaving the model itself unchanged. It lays out what kinds of parts can be changed and how candidate changes can be searched, checked, and kept; for example, a reusable tool might be created for one task and loaded on later, related tasks. It also derives conditions under which testing on new, independent examples can provide evidence that a revision helps rather than merely fitting past tests.

What it got

The main result is a set of conditional theoretical guarantees: expected performance differences are bounded under a stability assumption, accumulated regressions are bounded when per-round regression limits hold, and a fresh-validation rule can reject candidates unless the evidence supports improvement. The paper gives no single headline benchmark number; the guarantees depend on stated assumptions, including independent validation tasks.