What is the manager pattern in AI agents?
The manager pattern is a multi-agent design in which one central agent calls specialist agents as tools, keeps control of the conversation, and combines their results into the single reply the user sees.
OpenAI's "A practical guide to building agents" (April 2025) names two broad ways to split work between agents. In the manager pattern, a central agent coordinates specialists through tool calls. In the decentralised pattern, agents hand the conversation to one another as peers. The guide frames both as graphs: in the manager pattern the edges are tool calls, in the decentralised pattern they are handoffs.
The guide recommends the manager pattern when only one agent should control the workflow and talk to the user. Its example is a manager that receives a request to translate a phrase into Spanish, French and Italian, calls three translation agents as tools, and returns one combined answer. Google uses the name "agents as tools" for the same arrangement.
How it works in code
Wrap the specialist
In the OpenAI Agents SDK, agent.as_tool turns an agent into a tool, with a tool name and description the manager reads when deciding what to call.
Shape the input
By default the tool takes a single text input. Passing a schema lets the manager send structured arguments instead.
Shape the output
A custom output extractor can pull a JSON payload or a short answer out of the specialist's run before it reaches the manager.
Mind the history
The specialist's run does not inherit the manager's conversation by default. Sharing it means passing the same session to both, which should be a deliberate choice.
When to split at all
The same OpenAI guide says to maximise a single agent first, and to split only when instructions or tools become hard to manage. It gives two signs. Complex logic: prompts full of if-then-else branches that no longer scale. Tool overload: not the count of tools so much as their overlap, since some systems handle more than 15 distinct tools while others struggle with fewer than 10 overlapping ones.
Anthropic's October 2026 guidance on folding subagents into the main loop uses subagents as tools as the safer baseline when a system does keep several agents: the main agent owns the conversation, specialists return results in a shared response format, and the main agent verifies them. Full handoffs are kept for the cases where a specialist must talk to the user directly.
Where the manager pattern fails
An overloaded manager
The manager still has to pick the right specialist from tool descriptions. Vague or overlapping descriptions bring back the tool-selection problem the split was meant to fix.
Lost context
Specialists see only what the manager sends. A request that depends on an earlier turn fails if that detail is not passed along.
Double cost
Every specialist call is a full agent run, and the manager reads every result. Latency and tokens add up quickly for simple requests.
Weak synthesis
The final answer is only as good as the manager's merge. Evaluating each specialist alone misses errors introduced when results are combined.
How Motiontic's engineers apply it
Motiontic's engineers use the manager pattern when one voice must stay in front of the user while specialised capabilities sit behind it. Each specialist gets a narrow job, a clear tool description and a fixed output format, and the system is tested end to end as well as per specialist, because most failures show up in the hand-back.
Manager pattern compared with orchestrator-worker
| Manager pattern | Orchestrator-worker | |
|---|---|---|
| Specialists | A fixed set, each defined in advance as a tool | Workers spawned per input, often from one template |
| Who decides the split | The manager, choosing among known tools | The orchestrator, inventing subtasks at run time |
| Who talks to the user | The manager only | The orchestrator only |
| Typical use | Front-door assistants with distinct specialist skills | Breadth-first research or changes across many files |
| Main risk | Wrong specialist chosen, or context not passed on | Poor task splits and a high token bill |
Common questions
What is the difference between the manager pattern and handoffs?
In the manager pattern the central agent keeps control and gets each specialist's result back. In a handoff, control of the conversation moves to the specialist, which then answers the user itself.
What does "agents as tools" mean?
It means wrapping an agent so another agent can call it like a function: send an input, get a result back. The calling agent stays in charge of the conversation.
When should the manager pattern be used instead of one agent?
When one agent's prompt has become a tangle of conditional rules, or its tools overlap so much that it keeps picking the wrong one, and better tool descriptions have not fixed it.
Related
Agent handoff
One agent passes the conversation to another, which carries on from there.
Orchestrator-worker pattern
A lead agent plans, workers run in parallel with their own context, the lead merges.
OpenAI Agents SDK
OpenAI's code-first agent framework: a small set of primitives that run in a team's own app.
Single Agent vs Multi-Agent
One loop with good tools, compared with several agents splitting the work.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Sources
- OpenAI, A practical guide to building agents (PDF, April 2025)
- OpenAI Agents SDK docs, Tools (agents as tools)
- OpenAI Agents SDK docs, Agent orchestration
- Anthropic, When to fold subagents into the main agent loop (2026-10-22)
- Google Developers Blog, Architecting efficient context-aware multi-agent framework for production (2025-12-04)
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.