Back to Notes

Essay · July 11, 2026

The work should survive the conversation

What I want from coding agents: evidence, durable context, and work another person can actually resume.

Explore the public Workbench template

A coding-agent session can feel productive while it is happening. Files change, commands run, and the task moves forward. The harder question is what remains after the chat window closes.

That is the test I care about. Can another person understand the decision? Can a future session resume without reconstructing the whole story? Is there evidence that the result works, or only a confident summary saying that it does?

Leave a trail

Important reasoning should not live only in a conversation. Plans, decisions, and constraints belong somewhere durable enough to inspect later. A short file with the right facts is usually more valuable than a long transcript.

Make verification visible

The final answer is not the proof. Tests, rendered behavior, logs, and the actual diff are. I want an agent workflow to make that evidence easy to find and hard to hand-wave.

Design for the next handoff

Long-running work rarely ends in the same session where it began. A useful handoff says what changed, what was verified, what remains uncertain, and exactly where to continue. It should reduce guesswork instead of transferring it.

That is the idea behind Workbench. The system I use is private; the public template shares the repeatable parts. I do not want an agent that merely finishes a task. I want a way of working that leaves the result easier to understand than it found it.