The Blume Method
How we build software with coding agents: the work is organised around outcomes, evidence and one human decision.
Coding agents changed what is scarce. One person can have five agents writing at once, and the slow part is knowing what they built and whether it works. The Blume Method is how we work under that condition.
The problem
You ask three agents for three changes. Two hours later each one says it is done. You have three transcripts, three branches and three pull requests. To find out if the refund flow works, you check out the branch, start the app and click through it yourself.
Nothing says what is left before you can ship. “Done” is a sentence an agent wrote, not proof. And agents work in parallel, while your attention does not.
The answer
- The unit of work is a delivery: one change with an outcome a person asked for.
- The agent plans, writes down how the outcome will be checked, builds and runs the checks.
- Every check leaves evidence tied to the exact code revision.
- A person looks at the outcome with its evidence and decides.
The same afternoon: three deliveries, one waiting for you. You open it, see the refund flow as screenshots at the commit under review, and approve or write what should change.
Principles
- Deliver outcomes, not output
- “Done” means evidence
- A person decides
- Prefer a note over a block
- Phases follow reality
- Keep the plan true
Practices
- Agree on the outcome first
- Acceptance checks before code
- Reproduce, then fix
- Review the evidence first
- Run agents in parallel, decide one at a time
- Keep deliveries small
In Blume
Blume is a desktop app and a plugin for your coding agents. The plugin teaches each agent this process. The app shows you the result. Start with the start guide and the concepts, then deliveries and reviews.