meta-harness-open-thoughts-lite - Run one iteration of AgentHarness evolution.
Analyzes results and trajectories, proposes agent variants, implements them, and submits candidates for evaluation.
Tags
Updated: 2026-09-19Capabilities
Typical Inputs
Typical Outputs
What this skill does
- Analyze execution results and trajectories
- Propose candidate agent variants
- Implement candidate agent scaffolds
- Submit candidates for evaluation
Inputs
- frontier_val.json file
- evolution_summary.jsonl file
- Job and log trajectory outputs
- Agent source files in agents directory
- Prompt templates in prompt-templates directory
Outputs
- Candidate agent Python files
- New prompt template files
- Submitted candidate registration entries
Requirements
- HARBOR_MODEL environment variable
- harbor package with Terminus2 base class
- Python environment with litellm support
- submit_candidates tool