Multi-Agent Orchestration Patterns
Compare the canonical multi-agent orchestration patterns — sequential, group chat, supervisor, and human-in-the-loop — and learn which problem shape calls for which topology.
Multi-Agent Orchestration Patterns
Single agents are great for single problems. The moment a task crosses domains — read code, run tests, write a release note, ping the on-call — you need orchestration: a way of deciding which agent runs, in what order, and who closes the loop.
The exam expects you to recognise four canonical patterns and know when each is the right tool.
The four patterns at a glance
| Pattern | Control flow | Best for | | --- | --- | --- | | Sequential | Fixed pipeline, output of step n feeds step n+1 | Deterministic, repeatable processes | | Group chat (handoff) | Dynamic routing between peer agents | Conversational triage, expert escalation | | Supervisor | A coordinator dispatches specialists and aggregates results | Plan-and-decompose problems with a known goal | | Human-in-the-loop | A pause node awaits explicit user input | Approvals, clarifying questions, irreversible actions |
The Foundry workflow builder gives you templates for the first three out of the box and lets you splice in HITL nodes anywhere.
Exam tip: If the scenario describes a fixed ordered process, pick sequential. If it says "depending on the kind of request…", pick group chat or supervisor. If it says "before sending / merging / shipping / paying", the answer almost always includes a HITL node.
Sequence the pipeline
A research-then-summarise workflow is a textbook sequential pipeline. Below, drag the steps into the order they must run.
Quick check
Put it in order
Order the agents in a research-then-summarise pipeline that ends with a human approval.
- 1Critic agent reviews the draft
- 2Writer agent drafts the summary
- 3Extraction agent pulls structured facts
- 4Human approves before publish
- 5Retrieval agent gathers source documents
Where this shows up on the exam
GH-600 questions about coordination usually give you a scenario and four candidate topologies. The right answer is the simplest pattern that supports the determinism, branching, and approval needs of the scenario. When in doubt, default to sequential plus an HITL node for any irreversible step.
Key terms
- Sequential pattern
- A pipeline in which the output of one agent becomes the input of the next, in a fixed, deterministic order.
- Group chat (handoff) pattern
- Control is passed dynamically between agents based on context or routing rules; no agent owns the whole task.
- Supervisor (orchestrator) pattern
- A coordinator agent decides which specialist agent to invoke next and aggregates their outputs into a final answer.
- Human-in-the-loop step
- A workflow node that pauses execution and requires explicit user input or approval before continuing.
Common pitfalls
- Using a free-form group chat when the task is actually a deterministic pipeline — you trade reliability for nothing.
- Hard-coding a sequential pipeline for a problem with branching outcomes; the workflow can't represent the real decision tree.
- Adding a supervisor agent without giving it a clear stop condition, so it keeps re-dispatching specialists in a loop.
- Treating human-in-the-loop as optional polish — for irreversible actions, it is the orchestration pattern.