SafeEvolve: Harness-Policy Co-Evolution from Agent Experience for Safety Alignment
SafeEvolve is an experience-driven self-evolving framework for agent safety alignment that leverages safety experience from completed on-policy trajectories to drive a continual loop of harness-policy co-evolution. On the harness side, it converts trajectory-level safety evidence into bounded, component-level updates across safety prompt and hierarchical skills, yielding auditable and reversible harness artifacts. On the policy side, it follows a two-stage SFT-RL paradigm, where harness-use SFT bootstraps the policy to actively leverage evolved harness artifacts, and harness-augmented RL further shapes autonomous safety behaviors during multi-step exploration via verifier-decomposed rewards.
SafeEvolve proposes a self-evolving framework for aligning LLM-based agents with safety requirements by co-evolving the harness and policy from agent experience. It addresses safety risks in both final responses and multi-step execution trajectories, moving beyond isolated harness updates or policy optimization.
The framework introduces a closed-loop co-evolution mechanism: harness updates are bounded and component-level, derived from trajectory-level safety evidence, while policy optimization uses a two-stage SFT-RL approach with verifier-decomposed rewards to encourage safe multi-step exploration.
This approach signals a shift toward runtime-adaptive safety mechanisms for agentic systems, potentially reducing reliance on static alignment and enabling auditable, reversible safety artifacts in deployed agents.
For enterprises deploying LLM-based agents, SafeEvolve could lower safety incident rates and provide auditable safety updates, reducing compliance overhead and enabling safer autonomous operation in high-stakes environments.
Observable next signals include empirical evaluations of SafeEvolve on standard agent benchmarks, comparisons against isolated harness or policy baselines, and adoption of co-evolution techniques in commercial agent frameworks.