“Contractors you send a brief and never hear from until they're done; a task board with live workers you can glance at; and an actual engineering team that talks among itself.”
Level 1: Subagents
Subagents run inside your current session and report their results back to you. They can't talk to each other. They're best for repeatable, self-contained tasks (review, test, docs) where you hand off a clear brief and want it back. Think of them as contractors you send a spec.
Level 2: Agent View
Agent View is a full-screen interface over all your background sessions: dispatch new work, peek at any agent without interrupting it, and attach when one needs input. Sessions run under a supervisor process, so they survive you closing the terminal. Best for 3–10 independent tasks: a task board with live workers.
Level 3: Agent Teams
Agent Teams is the experimental, off-by-default mode where one lead agent coordinates teammates who can message each other and share a task list. That's the part that makes *dependent* work across files possible: a real team, not isolated workers. This is where the course is headed.
talk to survive coordinate
each other terminal dependencies
SUBAGENTS no no no repeatable tasks
AGENT VIEW no yes no 3-10 independent
AGENT TEAMS yes yes yes dependent, multi-file
Pick by the shape of the work, not by what sounds most powerful: a repeatable check is a subagent; a batch of unrelated jobs is Agent View; a feature whose parts depend on each other is an Agent Team.
In one line each
- Subagents: run in-session, report back, can't talk to each other. Best for repeatable tasks.
- Agent View: a dashboard over background sessions that survive terminal closure. Best for independent tasks.
- Agent Teams: experimental, off by default; a lead coordinates teammates that share a task list. Best for dependent, multi-file work.
Where to go next