harness-improvement-loop

AI harness improvement uses instrumented probes against the execution kernel to surface weaknesses in loop handling, context pressure, and tool usage. Collected runtime signals generate structured reports highlighting coverage gaps and measurable success criteria.

Automating harness improvements involves running a baseline probe suite across the kernel and reasoning strategies to collect runtime outputs. These outputs are aggregated into a detailed improvement report template that drives targeted fixes for context-management and tool usage.

Start a harness improvement pass by copying the REPORT-TEMPLATE.md file to your harness-reports directory and executing the probe suite. The suite outputs structured JSONL data and observations that populate the template with actionable findings.

The harness improvement loop integrates with agent-tdd to implement fixes while preserving historical reports and baselines. This ensures that templated improvements generated from probe data are executed without losing previous observability data.

Running instrumented harness probes requires the reactive-agents runtime, an LLM provider, and the effect framework to manage execution. Ollama is also required to provide local model inference for the probe suite.