Agent
Delivery Agents: build honestly, verify independently
The delivery agents turn a hardened specification into working software, one slice at a time. The key design choice is independence: no agent grades its own work.
The roles
- The Developer builds the slice and must produce real, working files.
- The Functional Tester checks it against the specification and never sees the Developer's reasoning.
- The Adversarial Tester tries to break the result when the risk of the slice calls for it.
- The Director runs on a different model and decides whether the slice is done.
Evidence over claims
The Director reads the raw output of every check, not the Developer's summary. If the two disagree, the raw output wins. The checks include static analysis, type checking, the test suite and a boot probe that proves the application actually starts.
Rules that stop shortcuts
- No empty code: every slice produces real files from end to end.
- Saved data is verified by restarting the application and reading it back.
- Audit and event records must reach storage, not only a log.
- An agent that cannot produce honest evidence says so, with a concrete reason, instead of inventing it.
See Marduk on your own idea.
Tell us what you want to build and we will show you the pipeline working on it.
Request a demo