Alibaba's coding-agent product Qoder launched Better Harness (Beta), a tool that assesses a coding agent's working environment and turns the lessons from each task into reusable assets. Rather than judging a single answer, it examines the whole 'harness' around the agent: whether goals and context are clear, the project is easy to run, permissions are controlled, verification is effective, delivery is safe, and both the team and the agent can learn from each task. The analysis runs in four steps — identify, spot gaps, apply the right fix (a rule, skill, hook, script, automation or human-approval gate), then recheck within a defined scope. To reduce bias, it asks three independent read-only subagents to interpret three kinds of evidence separately before combining them. It is available in the latest Qoder Desktop and as a Skill, with reports shown in Qoder Canvas, including which skills are actually used and which have never been called. Qoder says fixing these context-engineering issues also lowers token or credit consumption.
Our takeOur take: The jargon — prompt, context, harness, loop, graph engineering — arrives faster than anyone can learn it, and this tool quietly admits as much. The useful idea is small and sturdy: when the agent fails, build a mechanism so it cannot fail that way again. Grading the workspace, not the answer, is the right target. But a 'score your own setup' tool from the vendor selling the setup deserves skepticism — wait for independent evidence that the fixes cut real failures.