Self-Evolving Coding Agents via Harness-Policy Optimization
What this work does
Alternates harness search proposed by the current policy with RL updates under the selected harness, carrying hard tasks and failure evidence into subsequent rounds.
Connection to discovery
A concrete co-evolution mechanism for jointly improving discovery policies and the harness that elicits their trajectories.
Evidence
Inspected the H/P-step algorithm and coding-benchmark trajectories comparing policy-only, harness-only and alternating updates.
Scope and limitations
Evidence is from coding tasks under finite rounds and budgets; coordinated gains do not prove general recursive acceleration or scientific validity.
THE RESEARCH CONVERSATION
Discussion & reactions
Share a question, an insight, or a replication experience. Sign in with GitHub to join the conversation.