Loading…
Loading…
What it looks like to make an agent build to a written plan instead of a guess, and to put a second agent in front of the human reviewer.
Worked example · mechanism, not a customer story
Table of contents

An agent given a vague instruction produces plausible code that is not quite what you meant, and the gap only becomes visible in review. The cost lands on the human reviewer, who ends up reverse-engineering the intent from the diff. Generating code faster does not help if every change needs several rounds of correction.
The expensive part is not writing the code, it is discovering in review that the wrong thing was written.
Write the plan first. A feature becomes a spec with acceptance criteria, the files in scope, and what the tests should prove. Agents then work one scoped, reviewable change at a time rather than producing one sprawling diff.
Put a reviewing agent between the builder and the human. The builder implements; the reviewer checks the change against the spec and runs the suite. The human sees the change after it has already been held to the written intent.
Read the traceability. Because the spec and the change are linked, a change that keeps missing the mark tells you the spec was underspecified — which is information about your process, not just about the agent.
Review shifts from reconstructing intent to checking a change against intent that was written down first.
Again, mechanism rather than measurement. A spec-first workflow moves the thinking earlier and gives the reviewer something to check against; it does not guarantee fewer rounds, and any number we quoted here would be invented. Try it against your own review history if you want a figure.
Product updates and new worked examples as they are written. No customer stories until there are customers to quote.

Brand new and local-first. It extends the subscriptions you already pay for. Be among the first to put a command center behind them.
The Nerve Center for Agentic Engineering