Choose an orchestration pattern

01The goal

Choose between the main conversation, subagents, a dynamic workflow and an agent team by what a job needs and what it costs — and run one job as parallel subagents.

02The idea

Subagents, dynamic workflows and agent teams all split a job across workers. What differs is who holds the plan. With subagents, Claude decides turn by turn and each result comes back into your conversation. A dynamic workflow is a script Claude writes: the script holds the loop and the intermediate results, so it can run many agents, cross-check their findings, and rerun the same way. An agent team is a lead supervising teammates who message each other directly.

That last sentence is the whole discipline. The expensive pattern is the mis-chosen one: a team of investigators arguing over a typo fix, or a workflow fanning out to audit three files. We run the same economy on our delivery board: delegation is a decision with a price tag, not a default.

03The walkthrough

1. 01Pick three top-level folders in your repository — src, test and docs, or yours — and start claude.
2. 02Take one job: a newcomer’s summary of each folder. Ask the chart’s questions: one conversation would fill with file reads it will not need again — not the main conversation. The summaries need not talk to each other — no team. Three tasks, done once — subagents.
3. 03Ask for them the way the official parallel-research example does: "Research the src, test and docs folders in parallel using separate subagents, and give me two sentences on each." An interactive session runs subagents in the background by default, and a panel below the prompt shows a row for each while it works.
4. 04While they run, /tasks: each subagent listed with its state and its model; Enter on a row opens its transcript. Rows leave the panel when they finish successfully.
5. 05When the summaries land, /usage. On a Claude plan it attributes recent usage to skills, subagents, plugins and MCP servers — the number that makes the chart’s cost question real. Both the plan view and any per-session figure are estimates.
6. 06Name what would change the answer: the same summary for every folder, cross-checked and rerun before each release, is a workflow — the next lesson. Three investigators who must argue a flaky test’s root cause out between themselves are a team.

04Your turn

Four jobs for your repository. Choose a pattern for each — write the choices down with one line of reasoning — then run only the one that calls for subagents: check the three largest source files for errors that are caught and then ignored, one reviewer per file, findings only. (The other three: the same check on every file, verified and rerun before each release; a typo in one error message; three investigators testing different theories for a flaky test, messaging each other until they agree.)

Watch the prompts: a background subagent’s permission questions appear in your main session, and an answer scoped to the rest of the session applies to your main conversation too. Approve for that one call only, and ask reviewers to read, never to write. Keep the results short — every summary lands in your context window.