Working with agents

Creating an agent

Give it a job, an engine, and a folder, then a first task worth judging it on.

An agent keeps its own conversation, model, instructions, working folder, memory, tools and computer preference. Make a new one whenever the role, the access, or the project should be different.

#Make one

1. New in the sidebar, then New agent.
2. Give it a name and a clear purpose. The purpose is not decoration: it becomes the instructions the engine reads on every turn.
3. Pick its engine and model.
4. Choose a working folder if it should read or edit files.

The working folder matters more than it looks. It is the context the agent starts from and the place it is meant to act, which is also what keeps it away from everything else on your disk. A dedicated project folder is almost always the right answer.

#A first task worth judging it on

Ask for a diagnosis before you ask for a change:

Review this project, run the tests, and explain the three highest-impact
issues you find. Do not change any files yet. Point at the file and line
for each one.

Read the answer. If it understood the project, ask it to fix one. Splitting the reading from the writing keeps both steps small enough to check, and it is much easier to tell a good agent from a confident one this way.

#Shaping it

Agent settings holds the name, avatar, instructions, engine, model, working folder and computer. Instructions are the highest-leverage field: describe how you want the work done, not just what it is. Setting up a companion has the short version, and why you need no setup files.

Two agents can use different engines, folders and models at the same time. Nothing here is a global setting.

#Memory

Each agent keeps plain-text memory you can read, correct, or delete. It is not a hidden profile living on somebody's server; it is a file in your Pragma data. If an agent has learned something wrong about you or a project, open the memory and fix the sentence.

#Rooms and delegation

A room puts several agents in one conversation, each still running its own process. Everyone a message addresses answers at the same time, and the header shows who is replying. Use one when several specialists need the same brief. A chief-of-staff arrangement works well: one agent takes the request, hands out focused pieces, and summarises what came back.

Delegation is one agent asking another to handle a bounded piece. The delegated run keeps its own activity and reports completion, failure or cancellation back to the conversation that started it.

#Queueing and steering

Sending a message while an agent is working queues it rather than interrupting. Engines that support steering can be redirected mid-run, which is usually better than cancelling and starting again.