Verification
A feature is not finished because its builder says so. Automated checks decide, and a second agent tries to prove the builder wrong.
What it is
Verification has two parts. Automated checks (scripts) run and decide pass or fail from evidence. Then the breaking agent, a separate agent whose only job is to break the feature, tries to disprove its claims.
A check that produces no output has not passed. No output is not success.
How it works
Run the automated checks
Formatting, types, tests, builds, security and dependency checks run as scripts. The result is tied to the exact version of the code it checked.
Exercise the real system
The application is started and the changed behaviour is used, including desktop and phone width screenshots and an end to end run.
Break it
The breaking agent, which did not build the feature, tries unusual sequences, bad input, other paths, permission limits and failures.
Fix the root cause
Problems are fixed at the cause, a regression test is added, and the related flows are run again.
What it guarantees
- Pass or fail is decided by scripts, never by a model's opinion.
- A pass belongs to one version of the code. When the code changes, the pass is out of date and the checks run again.
- Every problem found becomes a permanent check, so the same defect gets harder to introduce.
What is not built yet
- No automated check exists in code yet.
- The breaking agent is a written method, not a running agent.
- Tracking whether a pass is out of date is not built.
Build with us as a design partner.
Methods is pre-alpha. A small number of teams will run it on real work while it is built, and shape what it becomes.