QoderAI/better-harness

What it solves

AI coding agents often move faster than the workflows surrounding them, leading to "fuzzy goals," improvised steps that cannot be reproduced, and a lack of proof that changes actually work. Better Harness addresses these system-level problems by analyzing the workflow around the code diff, rather than just the final output, to identify and prioritize gaps in the AI coding process.

How it works

Better Harness uses a feedforward-and-feedback loop to evaluate the "Agent Work Loop" across five dimensions:

1. Task Understanding: Checks if the agent knows the goal and definition of "done" using rules, specs, and design documents.
2. Controlled Execution: Ensures work follows repeatable paths via skills, commands, and sandbox boundaries.
3. Change Validation: Looks for evidence (tests, linters, diagnostics) that the change works.
4. Reliable Delivery: Verifies that quality checks and human reviews are not bypassed.
5. Learning Capture: Ensures lessons from one task benefit future tasks through memory and reusable skills.

It integrates as a plugin for various coding agents (such as Claude Code, Cursor, Codex, and GitHub Copilot) and generates evidence-backed reports (HTML or Markdown) that highlight prioritized findings and suggest scoped repair plans.

Who it’s for

Developers and teams using AI coding agents who want to move beyond reviewing diffs to optimizing their entire AI-assisted development lifecycle.

Highlights

• Multi-Host Support: Compatible with a wide range of agents including Claude Code, Cursor, Codex, GitHub Copilot, Qwen Code, and others.
• Evidence-Based Analysis: Explicitly tracks what is observed versus what is missing, avoiding unsupported scores or claims.
• Actionable Findings: Turns gaps into prioritized findings with impact assessments, expected outputs, and scoped fix plans.
• Visual Reporting: Provides HTML reports and a "Harness Inspector" for tracing delivery and agent activity.