# The method

How the Agent Workbench tools add up to one way of working with coding agents, one step per tool, with figures from the Lab write-ups.

## 1. Keep the backlog a schema

roadkeep owns the writes to the roadmap, the changelog and the rationale. An agent files and ships work through a command that checks the line's shape at insertion, and the Claude Code plugin refuses the hand-edit, so the backlog cannot drift when someone is in a hurry.

Tools: [roadkeep](https://alegauss.github.io/products/roadkeep/)

## 2. See what every project owes

roadkeep-gui finds every governed checkout on the machine and shows what each has shipped and still owes. It files or ships a line through the roadkeep command, and hands a task to Claude Code as a session it starts, claims and shows while it works.

Tools: [roadkeep-gui](https://alegauss.github.io/products/roadkeep-gui/)

## 3. Let the diff write the commit

commitclerk reads the staged diff, asks an LLM for a Conventional Commits message, shows it to you and commits only once you agree.

Tools: [commitclerk](https://alegauss.github.io/products/commitclerk/)

## 4. Test what was observed

winwright drives the Windows app from a test and reports four verdicts, so a check that could not run is never filed as one that passed.

Tools: [winwright](https://alegauss.github.io/products/winwright/)

## 5. Give the agent a domain it can check

polyweave declares what a 3D asset must satisfy and searches for parameters that do, asserting every render; Piano's plugin writes, checks and keeps scores before the app plays them. Same discipline: a declared target, checked, recorded.

Tools: [polyweave](https://alegauss.github.io/products/polyweave/), [Piano](https://alegauss.github.io/products/piano/)

## 6. Keep the usage in view

Claude Code Tray puts the rate-limit percentage in the tray, read from the session files already on disk.

Tools: [Claude Code Tray](https://alegauss.github.io/products/claude-tray/)

## Measured, not claimed

- 530 h of attributed time, across 556 sessions in 59 days — “59 calendar days, 556 sessions, 31 projects, 530 hours of attributed time” (https://github.com/alegauss/claude-code-time-study#result-on-the-measured-sample)
- 11.2% of active time went to authoring the feature — “feature authoring is 11.2%, less than half the expected share” (https://github.com/alegauss/claude-code-time-study#result-on-the-measured-sample)
- 27.3% went to writing and running tests — “Tests (writing + running) | 124 | 27.3%” (https://github.com/alegauss/claude-code-time-study#result-on-the-measured-sample)
- 1,864 tasks measured, over 79,377 requests in 32 active days — “The current snapshot covers 1,864 tasks and 79,377 requests over 32 active days.” (https://github.com/alegauss/claude-code-usage#measuring-instead-of-assuming)
- $777 for the most expensive five-hour window on record, at list prices — “the most expensive five-hour window on record cost $777 at list prices with work uninterrupted” (https://github.com/alegauss/claude-code-usage#assumptions-stated-so-you-can-disagree-with-them)