Early Access. Commands and behavior can change between releases.
Orchestrator
What it does. The Orchestrator is the coordinator. It reads your project board, routes work to the other agents, dispatches them and tracks their progress. It never builds anything itself. When it’s used. Every session starts here. Use it to pick up planned work, send it out and check on it.Architect
What it does. The Architect turns what you want into requirements, a design and a set of tasks. It splits the work so several agents can build in parallel. When it’s used. Before anything is built. Use it when a feature needs a plan, or when you want a design decision made first.Developer
What it does. The Developer builds the change. It writes the tests first, then the code, and checks that both pass. It can also review code against your project’s rules. When it’s used. Once a task is planned. The Orchestrator usually sends it the work for you.LLM-Whisperer
What it does. The LLM-Whisperer handles AI work: prompts, LLM features and agents. It measures quality with evals, then improves the prompt or agent one change at a time and keeps only what scores better. When it’s used. Whenever the change is about how a model behaves, not ordinary code.System agents and your own agents
System agents are read-only. In the/agents manager you can see them and their status, but you
can’t edit, export, delete or disable them. Open one and you see a notice that its definition is
read-only. You can still dispatch a system agent and steer it while it runs.
Your own agents are separate. Give them their own names, and you edit and manage them as usual.
To tune how system agents work in your project, don’t edit them. Put build commands, architecture
constraints and team conventions in project context and the
rules /onboard sets up.
Orchestrator and agents
How the roles work together, locally and in the Factory.