Glossary

Agent. A program that uses a model to plan, call tools and act.

Dark factory. A plant that runs with no people on the floor. In software, a pipeline that turns a specification into tested code with little human involvement.

Holdout set. Tests or scenarios the building agent cannot see. They judge the result from outside. (Order 005)

Oracle. The check that says whether work is correct. In this template, the test suite named in FACTORY.md.

Production order. A GitHub issue whose body is the spec. One order runs to one pushed commit.

Converge. Run the oracle until it is green. The iteration count goes in METRICS.md.

Autonomy contract. The three written lists: what the agent decides, decides and flags, and stops and asks.

Skill. A rule file or tool bundle an agent loads. It is part of the agent’s toolbox and a supply-chain risk. (Order 011)

Spec-driven development. A method where a written specification is the central artifact and code is the output. GitHub Spec Kit is one toolkit. (Order 002)

Prompt injection. Attacker text that an agent reads as an instruction.

Control plane. The part of a system that enforces policy. The history shows it moving out of the agent. (Order 043)