A linter for your AI agent's workflow

What it does

Better Harness is a plugin-backed analyzer that treats the process around AI coding agents as a first-class artifact. It gathers project and session evidence from hosts like Claude Code, Codex, and Cursor, then evaluates the workflow against five dimensions—task understanding, controlled execution, change validation, reliable delivery, and learning capture. The result is a prioritized report where each finding carries impact, a scoped repair plan, and acceptance checks; if evidence is missing, the gap stays visible instead of being papered over.

The interesting bit

The project applies a feedforward-and-feedback loop to agent engineering: AGENTS.md and specs steer the agent before it acts, while linters, tests, and hooks observe after. It is deliberately modest about its own conclusions—trends are recorded, but causal proof of improvement requires a comparable later run, not just a passing check.

Key highlights

• Scores the Agent Work Loop across five dimensions, from goal clarity to whether the next task learns from this one.
• Emits self-contained HTML or Markdown reports with evidence-bounded findings and scoped repair actions.
• Ships host adapters for Claude Code, Codex, Qoder, Cursor, and GitHub Copilot CLI.
• Leaves missing or partial evidence explicit rather than inventing unsupported scores.
• Open-sources the evaluation model, engineering practices, and runnable implementation as three connected layers.

Caveats

• Cursor support currently requires source-local loading; the plugin has not reached that marketplace yet.
• Adapters for Pi and WorkBuddy remain support entries rather than fully verified quickstart hosts.
• Historical trend views show recorded patterns, not causal proof that the loop itself improved.

Verdict

Teams using AI agents who keep hitting the same workflow friction should try this; if you only care about faster diffs, it will feel like bureaucratic overhead.