Impressive paper on building recursive self-improving agent harnesses.
It's rich with great insights on building effective agent harnesses.
If you maintain an agent harness, this paper names three defects in how harnesses get improved and the fixes.
First, evolving a harness against the evaluation benchmark makes reusable improvements impossible to tell apart from benchmark fitting.
Instead, ModularRSI evolves on tasks disjoint from the benchmark.
Second, updating from a single trajectory confuses a systematic harness deficiency with one task's reasoning details, which produces changes that fail on unseen tasks.
ModularRSI contrasts successful against failed trajectories for the same task, then aggregates across tasks to find recurring behavioral deficiencies.
Third, a monolithic harness makes it hard to localize a recurring problem, and optimizing the whole thing entangles unrelated mechanisms, so no change can be attributed.
ModularRSI splits the evolvable harness into five functional modules that evolve separately. Agent Loop, Tool Use, Observation Management, Context Management, and Task Completion Detection.