Better Harness is an open-source tool, published by QoderAI under the MIT license, for analysing and improving the workflow around AI coding agents. Its README summarises the idea as “Delegate coding to agents. Improve the loop around them”: the tool runs through the coding agent a developer already uses, gathers project evidence (and, on hosts where it is supported, session evidence), and turns it into prioritized improvements and verifiable next steps. The repository’s own description calls it a Harness Engineering platform for coding agents.
The problem it addresses is framed as one that reviewing the final diff cannot see. The README lists five ways the workflow around an agent goes wrong — fuzzy goals, so that the agent confidently solves the wrong problem; improvised steps that nobody can reproduce; “it works” claims without proof; review and delivery checks bypassed for speed; and lessons lost, so the same friction returns on the next task — and positions Better Harness as analysing those system-level problems instead.
Capabilities
Better Harness organises its analysis around a feedforward-and-feedback loop, citing Martin Fowler’s
site for the framing (see Guides and sensors): feedforward guides such as AGENTS.md,
specs, skills and acceptance criteria steer the agent before it acts, and feedback sensors such as
linters, tests, hooks and evaluation agents observe the results and help it self-correct. Across that
loop it evaluates five dimensions of delivery that it calls the Agent Work Loop:
| Dimension | The question it asks |
|---|---|
| Task Understanding | Does the agent know the goal and what “done” means? |
| Controlled Execution | Is the work on supported, repeatable paths? |
| Change Validation | Is there evidence the change actually works? |
| Reliable Delivery | Does AI speed bypass quality checks or acceptance? |
| Learning Capture | Does the next task benefit from this one? |
Running the /better-harness workflow establishes a task-bounded baseline and produces a report
combining a five-part overview, prioritized findings, the agent assets it detected and an evidence
brief. Each finding is tied to its evidence and carries an impact, an expected output, a scoped repair
and acceptance checks; a repair action drafts a scoped fix plan for review. Depending on the host, the
report is a host-native Canvas (Qoder and Cursor) or a self-contained HTML file with paired Markdown
(Claude Code, Codex, Qwen Code, GitHub Copilot and Kimi Code). Once comparable reports exist over
time, a history view shows how the five dimensions move, which the project presents as recorded trends
rather than causal proof of improvement. A separate interactive Harness Inspector traces delivery from
product intent through agent activity, sessions, files and commits in a read-only workspace.
The project stresses that missing evidence stays explicit. Unobserved behaviour is reported as such rather than turned into an unsupported score, and the README draws the boundary this way: configured assets can establish that a mechanism exists, but only linked task evidence can establish that it was used or improved an outcome. Internally, three independent evidence agents collect from separate domains, and a single lead agent unifies their results.
What the repository opens is described as three connected layers: engineering-practice guidance
(covering session evidence, project harness, agent customisation and
Loop Engineering), the Agent Work Loop evaluation model, and the runnable
implementation — the canonical /better-harness workflow packaged as an
Agent Skills skill, together with evidence collectors, analysers, renderers and thin
per-host adapters. A standalone CLI can inspect local installation evidence for every host and build
install, update or removal plans for a host without executing them.
Adoption & Ecosystem
Better Harness does not use one entry point across hosts; it is installed separately per coding agent, mostly as a plugin through each host’s marketplace mechanism. The README gives inline setup for Claude Code, Codex Desktop and the Codex CLI (OpenAI Codex), Qoder, Cursor and the GitHub Copilot CLI, and points to its documentation for Qwen Code, Pi, Kimi Code, WorkBuddy and Grok. It is built into the Qoder desktop app, where it needs no installation and is also reachable from a sidebar entry in Qoder’s Quest view, and a Qoder CLI installed alongside the desktop app picks it up as well. The Cursor plugin is not published to a marketplace, so the tool reports its Cursor installation plan as unavailable rather than emitting an unverified command. Session evidence depends on what each host records — for GitHub Copilot, for instance, it reads Copilot CLI transcripts, while noting that Copilot records no per-response token usage and that VS Code Copilot Chat has no supported durable transcript.