Early Access. Commands and behavior can change between releases.
Agent tool, the agent runs in
the background, and the session keeps going. Messages flow both ways: the session can steer a
running agent and read its result, and agents send findings back. You watch and steer all of it
from two places: the fleet list and the agent viewer.
In the full Architech session (architech), the orchestrator starts sub-agents when work needs a
specialist or can run in parallel. In agent mode (architech agent), your agent has the same tools,
and you can ask it for parallel work directly.
The /agents roster includes developer, llm-whisperer, architect, general-purpose
and explore. See orchestrator and agents for their responsibilities, and
system agents for which definitions are read-only.
The fleet list
Dispatched agents appear under your prompt:+ N more, never dropped. An agent keeps its number while it is in the list: removing one
never renumbers the others. Once an agent leaves the list, a new agent can take its number. Check
the row before you open a number you used earlier.
Finished agents fade from the list after one second. If you have an agent open in the viewer, its
final transcript stays readable after its row is gone. The list hides while a picker such as
/model is open.
The agent viewer
Open an agent and its conversation replaces the chat; it is not a split pane. The transcript looks like the main session: tool calls are cards, and markdown, diffs and errors render the same way. It updates live while the agent works. The input box at the bottom works like the main chat, and whatever you type steers the agent you are watching. If a message can’t reach the agent, your text stays in the box with the reason; you can’t steer an agent that has finished.Esc closes the viewer and returns you to the main
conversation; the agent keeps running.
Resume a sub-agent
A sub-agent’s result ends with its id:- the full id from the end of its result;
- the short ref shown in the agent list;
- a unique prefix of the id, at least four characters long.
Which model a sub-agent runs on
A sub-agent runs on the model its definition or harness model setting names. Otherwise it runs on the model of the session that started it. Architech can also ask for a size instead of a model:large, medium or small.
If none of a requested size’s models is configured on your machine, or the model its definition
names isn’t configured, the sub-agent still runs, on the session’s own model. You and the model
both see a warning naming the model it runs on instead. If a request names a specific model that
isn’t configured, Architech refuses it; it never swaps models silently.
Failures report as failures
If an agent never reached the model (no tokens, no tools, no output), you get an error with the reason (model availability, provider credential, rate limits), not a finished agent with an empty transcript. A sub-agent that stops for inactivity reports a failure and how long it was inactive. A resumed sub-agent never returns its previous result as if it were new.Questions from background sessions
An Architech session can start another session in the background. When that background session asks a question, the question goes to the session that started it:
A detached background session waits 10 minutes by default when a person answers, and 5 minutes
when an agent answers. Teammates in panes work differently: their questions wait for you with no
timeout, and a model never answers them.
Agent teams in terminal panes
Agent teams are experimental and off by default. With teams on, the orchestrator can start full interactive teammates in iTerm2 panes, not only detached background sessions.
Teams need iTerm2 with its Python API enabled. tmux panes are not supported. If no pane backend is
available, a teammate can run detached, and the session tells you why. A teammate can’t start
another team.
Orchestrator and agents
The software-engineering roles behind the fleet.