StarHarness is a framework that evolves the agent harness (prompts, tools, skills, subagent structure, etc.) while keeping model weights fixed. It uses a stratified search over task failure modes, separating search tasks from selection tasks and reserving held-out tasks for evaluation. Across three enterprise benchmarks—ITBench SRE, EnterpriseOps-Gym ITSM, and AutomationBench Finance— improved performance by 20–35 percentage points after 4–12 accepted changes per environment. The gains persisted on tasks excluded from evolution and transferred across GPT and Qwen models without re-evolution. This is interesting because it addresses persistent model-environment mismatch in tool-rich enterprise tasks through practical, code-level harness edits that can be tested and reverted.
Evolved agent harnesses lift enterprise performance without retraining models
StarHarness adapts prompts, tools, skills and agent-loop structure to specific environments while keeping model weights fixed.
Top University
Esakkivel Esakkiraja · Denis Akhiyarov · Vikas Yadav · Sai Rajeswar · Patrice Bechard · Sridhar Nemala · +1 more
ServiceNow · Mila · Université de Montréal
Research Digest··2 min read
Thread:Agent Harness Optimization
In one line
StarHarness evolves environment-specific agent harnesses, improving benchmark performance by 20-35 percentage points.
Before you read this
Two ideas this paper builds onRead it as
The authors developed a search framework that iteratively modifies an agent’s surrounding harness using tasks selected according to baseline failure patterns.
StarHarness evolves environment-specific components around a fixed language model, including task framing, prompts, tool interfaces, skills, MCP-backed providers, subagent organization and agent-loop configuration.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.