How is this different from orchestration?

Orchestration is the general coordination of steps and agents; a manager agent is a specific pattern where an agent does the coordinating.

Does the manager agent need to be the strongest model?

Not necessarily. Its strength comes from decomposition and acceptance criteria, not from being the cleverest component.

What is the main cost risk?

Unbounded delegation: every subtask is model calls, so plans need step and spend caps.

AI Glossary / Agents and automation / Intermediate

Manager agent

Definition

A manager agent is an agent that coordinates other agents: it breaks a task into parts, assigns them, checks the results and combines them, the way a team lead coordinates a team.

What it means

In multi-agent designs, work is split between specialized agents, and someone must run the project. The manager agent decomposes the goal, delegates to the specialists, monitors progress, resolves conflicts between their outputs and assembles the final result.

The pattern is attractive because specialization works: a research agent, a drafting agent and a checking agent can each be simpler and better tested than one agent doing everything. The manager turns the collection into a system.

It is also where complexity concentrates. Every delegation is a handoff with the usual risks: lost context, misunderstood instructions, duplicated work. A manager agent needs the same things a human manager needs: clear briefs, status visibility and a quality bar.

The checking role matters most. A manager that accepts whatever its specialists return is a mail forwarder, not a manager. Designs that work include acceptance criteria for each delegated piece and a plan for what happens when a piece fails.

For a business, the manager-agent idea is familiar territory: the disciplines that make human delegation work, briefs, checkpoints and acceptance review, are the disciplines these systems need.

How it works

A typical run has five stages. The manager interprets the goal and decomposes it into subtasks. It assigns each subtask to a specialist with a brief and the context it needs. Specialists execute and report back. The manager validates each result against acceptance criteria, retrying or escalating failures. Finally it synthesizes the accepted parts into the deliverable.

State management holds it together: a shared record of subtask status, intermediate results and decisions, so the manager can recover from failures without restarting everything.

Cost control lives at the manager level too. Every delegation is model calls and tool calls; the manager's plan decides how much compute a task consumes, so budgets and step caps attach to its plan.

Evaluation focuses on the seams: how often subtask results fail acceptance, how often the synthesis contradicts a part, and how often a human has to intervene.

Three practical examples

Research report: the manager assigns source gathering to one agent, extraction to another and drafting to a third, then checks citations before assembling the brief.
Month-end pack: one agent pulls ledger data, one drafts variance commentary, one validates figures, and the manager sequences them to meet the deadline.
Onboarding: document checking, account provisioning and welcome drafting run as delegated subtasks, with the manager escalating any failed validation.

Think of it this way

Think of a general contractor building a house. The contractor does not pour concrete or wire the kitchen; the contractor plans the sequence, hires the trades, inspects each stage and refuses to pay for failed work.

The limit: a contractor brings judgment the contract never wrote down. A manager agent inspects against its stated criteria only, so the acceptance criteria carry the quality standard that a contractor's eye would supply.

Worked example

delegating a market brief

Suppose a manager agent is asked for a two-page brief on a new market. It decomposes the work: gather sources, extract key facts with citations, draft the brief, verify every claim against the gathered sources.

Each subtask gets a brief with acceptance criteria: sources must be dated, facts must carry a citation, claims must trace to the gathered material. The verification agent rejects two claims that the drafter included from its own prior knowledge rather than the sources.

The manager sends those sections back with the rejection reasons, accepts the corrected version, and assembles the brief with a source appendix. The scenario and counts are invented; the criteria-driven loop is the working pattern.

The human reader still reviews the final brief. The manager's job was to make that review fast by making every claim traceable.

Illustrative case study

A fictional agency builds a manager agent over five specialists for client reports. Early runs produce polished reports whose sections contradict each other, because each specialist worked from a slightly different summary of the client's question.

The fix is a shared, manager-maintained statement of the task that every specialist receives verbatim, plus a consistency check at synthesis. Contradictions drop sharply.

The invented lesson: in managed agent teams, most quality failures are communication failures. The shared brief is the manager's most important artifact.

Common misconceptions and mistakes

More specialists mean better results. Each added agent adds handoffs; add a specialist only when its subtask fails as part of a larger agent's job.
The manager understands the task best. It coordinates against criteria; it can assemble fluent nonsense if its acceptance checks are weak.
Delegation removes the need for humans. People still set the goal, review the deliverable and audit samples of the managed process.

Business use

Manager-agent designs suit multi-part deliverables: research packs, recurring reports, onboarding flows and any task where subtasks benefit from separate testing.

Adopt incrementally. Start with one manager and two specialists, write acceptance criteria for each subtask, and scale the team only where the acceptance failures show a real need.

Check your understanding

A managed agent team produces a report whose sections contradict each other. What is the most likely cause and the first fix?

Answer: Specialists worked from inconsistent briefs; the fix is a single manager-maintained task statement given verbatim to every specialist plus a consistency check at synthesis.

Related terms

Learn next

AI Agents and Automation for Small Business teaches delegation-style workflows with approvals and error alerts: https://aiforbusinessandlife.com/academy/ai-agents-automation-small-business. [C4]

Sources

[S2] Google, Machine Learning Glossary. [S5] Anthropic, Building effective agents. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Managed private cloud Next: Mandatory access control