Build agents

There are three ways to put an agent to work. Each ends in the same place: a card in your deployment’s catalogue, released through the same gate.

Way	When to use it	State
From the library	A ready-made agent fits the job	Built: you request it in LightHouse; it reaches your deployment through a reviewed merge
With the builder	You want to describe the job in plain words	Coming
In code	Your engineers write the agent, or change one	Built: a card and a Python function, run by the engine’s orchestrator

What every agent has

• A card: its tools, its budget, its effect and its autonomy. See Agents.
• Evaluations: a golden set run before every release, and again against what is released. See Evaluations.
• Approvals: a person signs off any output with an effect outside the run, unless a standing pre-approval covers it. See Approvals.
• Versions: only a version that passed the release gate can be made current, and a breach rolls back to the last version that passed. See Versioning and rollback.
• A run map: every step of every run, with its model, cost and sign-offs. See The run map.