Google and CMU Paper: Multi-Objective Self-Evolution Keeps Agent Security Gains
rohanpaul_ai · x · 2026-10-15
A new Google and Carnegie Mellon paper, "Harmony Harness: Multi-Objective Self-Evolution of Agent Harnesses," finds that self-improving agents retain security gains far better when security is evolved as a separate objective.
Problem: Agent harnesses — the prompts, tools and loops around a model — can now rewrite themselves from execution feedback. But most such systems chase a single score (e.g. test pass rate) and evolve the harness as a monolithic artifact, so security gains mostly vanish on new tasks and transferable design can't be separated from environment-specific adaptation.
Method: Harmony Harness formulates harness evolution as multi-objective optimization, using decomposition-based evolutionary search to maintain solutions across the Pareto frontier and avoid hand-designed structural diversity metrics. Factorized evolution operators separately refine general-purpose harness mechanisms and benchmark-specific adaptations, combined at inference time via a dynamic composition workflow.
Results: Across SecRepoBench and BaxBench, Harmony Harness consistently beats strong harness-evolution baselines, with gains exceeding 10% on both. The authors conclude that explicitly modeling objective trade-offs and harness structure yields stronger, more transferable agent evolution than single-objective, monolithic approaches.
More from coding & agent
- Running a multi-model UI design arena: Paseo orchestrates four design drafts per model — vista8 · 2026-10-16
- Cortex A/B-tests coding-agent skills and rules by replaying developer corrections before keeping them — Samuele Giampieri · 2026-10-16
- Tool calls are the hidden bottleneck killing AI coding speed — Crabbox and 5 fixes — steipete · 2026-10-16
- Agentic sandbox WorldBox doubles GPT-6 Astra's MazeBench score, 23 to 61 gems — ZhitingHu · 2026-10-16
- Two prompts: Gemini 4 Argon turns a single-player game into full multiplayer — chongdashu · 2026-10-16
- Anthropic and Gray Swan launch $300,000 indirect prompt injection red-teaming contest — bcherny · 2026-10-16