Build an agent

You have a working agent. Each page here adds one capability and the guarantee it buys. Stop wherever you have what you need.

1. Add a second agent. A role says which fields an agent may write. Adding one changes no agent you already wrote.
2. Handling rejections. Someone changes a field while your agent is deciding. The write is refused and names what moved, so no agent acts on a decision that has been replaced.
3. The agent loop. Many records and many copies, claiming work so exactly one acts on each record, with no leader handing it out.
4. Surviving a crash. An agent dies holding a task. Its replacement takes over and resumes from the checkpoint, knowing what changed meanwhile.
5. The audit trail. What the record keeps, including the writes it refused, and how to check your agents behaved.
6. Design a class. Fields, modes, roles and evidence for a problem of your own.
7. The incident fleet. All of it at once: four agents on one record.

The same agent in other clients: Go and any MCP client. The same agent inside someone else’s framework: Recipes.

Every output on these pages comes from running the code shown. Most use the gateway from Install. The ones that need another class say so and show how to start it.