Every session so far has been you and one agent, taking turns. That breaks down once a job outgrows one conversation. The fix is a different job for you: stop typing every step, start deciding what gets done, by which worker.
When you run more than one agent you stop being one more person typing and become the orchestrator: the one who breaks a job into pieces, decides who does each piece, and combines the results. The pieces run as subagents, with their own context window.
Not every piece deserves the same model. A mechanical rename does not need the model you would trust with an architecture call. Cheap for mechanical work, mid for ordinary content and code, top reserved for judgment you would regret getting wrong.
Pin a whole session to one named subagent instead of re-describing it every time. Or force every subagent onto one model regardless of what it asks for. Either way, the one sentence dispatch pattern still starts the actual work.
For work that needs pieces to compare notes, not just report back, agent teams let a lead spawn teammates who message each other directly and share a task list. It is experimental and off by default, one environment flag away.
Start with three to five workers, whether subagents or teammates. Coordination overhead grows with every worker added, and returns shrink past a point. Fifteen independent pieces want three to five workers taking several each, not fifteen at once.
Every extra worker multiplies token spend, it does not just add to it. Five subagents burn roughly five agents worth of tokens, real money on a usage based plan. Spend the top model where a cheap one getting it wrong would embarrass you.
You are done when you can name which model belongs on which kind of work, you have dispatched at least three pieces with one sentence, and you can tell a job that wants plain subagents from one that wants teammates talking to each other.
That is the operating model: one orchestrator, many hands, each priced to the job. Next, session two makes sure every piece you hand off carries a brief no subagent can misread, because a blank context window only builds what you actually wrote down.
Every session so far has been you and one agent, talking in turns. That works until the job is bigger than one conversation: a dozen pages to write, a dozen small features to build, a research sweep against a real deadline. The fix is not a smarter agent, it is a different job for you. You stop being the one who types every step, and become the one who decides what gets done, in what order, by which kind of worker. That is the whole shift this course teaches: one orchestrator, many hands.
This is session 1 of Run an AI Team, free to everyone. By the end you will have picked a real batch of work, matched a model to each piece of it, and dispatched more than one agent from a single sentence.
When you run more than one agent, you stop being "one more person typing" and become the orchestrator: the one who breaks a job into pieces, decides who does each piece, and combines the results. The pieces themselves run as subagents, helper sessions with their own isolated context window and their own tools, spawned from your main conversation and reporting back to it. You already met this shape in session 5 of Your AI Coding Agent, one dispatch at a time. This course runs it as a standing operation: many pieces, spread across hours or days, each tracked until it lands.
Not every piece deserves the same model. A mechanical rename across forty files does not need the model you would trust with an architecture decision, and paying for that is money spent on nothing. Claude Code subagents take a model field right in their definition:
---
name: renamer
description: Mechanical rename across the codebase
model: haiku
---
model accepts sonnet, opus, haiku, fable, a full model ID, or inherit to match whatever the main conversation is running. A simple rule that holds up in practice: cheap model for mechanical, repetitive work; mid-tier for ordinary content and code; top-tier reserved for design, architecture, and anything you would regret a cheap model getting wrong. To force every subagent in a session onto one model regardless of what each one asks for, set an environment variable instead of editing every file:
CLAUDE_CODE_SUBAGENT_MODEL=haiku
The one-sentence dispatch pattern from Your AI Coding Agent still works here: "Use subagents to do these in parallel: ..." followed by the list. For a standing team, name a subagent once instead of re-describing it every time:
claude --agent code-reviewer
That pins the whole session to one named subagent, the same one every time you ask for review, instead of retyping "use the code-reviewer subagent" on every request.
For work that genuinely needs the pieces to compare notes, not just report back, Claude Code has agent teams: a lead session that spawns teammates who message each other directly and share a task list, instead of everything routing back through you. It is experimental and off by default:
{
"env": { "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1" }
}
With that set, a prompt like "spawn three teammates to draft the grant sections in parallel, one per section" gets you a lead that assigns work, teammates that claim tasks and message each other, and a synthesis at the end. Reach for a team when the pieces benefit from disagreeing with each other; reach for plain subagents when each piece just needs to finish and report back.
Start with three to five workers, whether subagents or teammates. Coordination overhead grows with every worker added, and returns shrink past a certain point: three focused workers usually beat five scattered ones. If a job breaks into fifteen independent pieces, three workers taking five pieces each is a better starting point than fifteen workers at once.
Every extra worker multiplies token spend, it does not just add to it: five subagents burn roughly five agents' worth of tokens. That is real money on a usage-based plan. Spend the top-tier model on the piece that would embarrass you if a cheap model got it wrong, and let the cheap model earn its keep on everything mechanical.
You can name, out loud, which model belongs on which kind of work, you have dispatched at least three pieces with one sentence and used the combined result, and you can tell the difference between a job that wants ordinary subagents and one that wants teammates who talk to each other. Session 2 gives every piece a brief good enough that a worker cannot misread it.
Sources: https://code.claude.com/docs/en/sub-agents, https://code.claude.com/docs/en/agent-teams
Template included: Get the role script (who does what), the model-tier decision card, and the team-size rule of thumb as a one-page reference. Find it in your member dashboard under Templates.
Short on time? Agentic Quarks will implement this end to end, wired into your stack, and hand you the keys.
One orchestrator, many hands, and briefs no agent can misread
6 sessions, 190 minutes, 1 free.
One payment, no subscription, and a 14-day money back you trigger yourself from your account page. This course on its own is $99.