An agent team is a coordination model

An agent team is organized work among several agents with a lead, separate contexts, a task list, handoffs, and a human decision point. Several open sessions or independent subagent calls do not become a team automatically.

A team can research different parts of a difficult task in parallel, but information is not shared automatically.

RoleResponsibility
LeadDistributes work and synthesizes results
TeammateCompletes a narrow research or execution step
Shared task listShows status and ownership
HandoffTransfers findings, evidence, risks, and links
HumanApproves risky changes and final outcome

Read-only team example

Parallel research streams
backend-researcher: business rules
ui-researcher: inbox state and rendering
qa-researcher: tests and gaps

lead: synthesize evidence
owner: decide next step

Team workflow

  1. Owner defines goal, boundaries, and risks.
  2. Lead assigns independent read-only research.
  3. Each teammate returns evidence, files, and risks.
  4. Lead identifies conflicts and builds a shared view.
  5. Human approves the next step.
  6. One executor makes changes.
  7. Diff, checks, and outcome receive separate review.

Human checkpoint

Keep a human in the loop before broad edits, public API changes, payment or authorization changes, merge, release, or rollback. A team collects facts and hypotheses; it does not take product responsibility.

Canonical lesson source · JavaRush

Open JavaRush material