Supervisor Agents
A supervisor agent coordinates a team of specialized sub-agents, routing each user request to the right specialist. Tridan AI builds this multi-agent orchestration on LangGraph.
What is a supervisor agent?
Instead of one assistant trying to know everything, a supervisor sits in front of several focused sub-agents — for example a billing specialist, a technical specialist, and an order-status specialist. The supervisor reads each incoming request, sends it to the right specialist, and returns a single coherent answer to the user.
The supervisor also controls what each agent can see and when. In phase-based workflows (such as a structured sales call), each phase's agent only gets the context it needs, so agents never re-ask questions that were already answered earlier in the conversation.
Single assistant or supervisor team?
Swipe sideways to see the full table
| Use a single assistant when | Use a supervisor team when |
|---|---|
| One topic area with one knowledge base | Requests span distinct domains (billing, technical, orders) |
| A single set of instructions stays short and unambiguous | Instructions for one agent would grow long and contradictory |
| Conversations are free-form Q&A | Conversations follow phases that need strict context control |
| Few or no external tools | Different tasks need different tools or knowledge bases |
How routing works
- 1
The supervisor reads the request
Every user message goes to the supervisor first. It interprets the intent and decides which specialist is best placed to handle it.
- 2
The right specialist takes over
The supervisor routes the conversation to that sub-agent along with only the context that agent is allowed to see for the current phase.
- 3
The specialist does the work
Each sub-agent uses its own instructions, knowledge base, and tools — a billing agent can query invoices while a technical agent searches product docs.
- 4
Results flow back through the supervisor
The supervisor collects the specialist's answer, keeps the conversation coherent, and decides whether to respond, route to another agent, or move to the next phase.
Example: a support team
A typical customer-support setup pairs one supervisor with three specialists. Specialists can call custom tools to look up live data such as orders or invoices.
Swipe sideways to see the full table
| Agent | Handles | Knowledge & tools |
|---|---|---|
| Support Supervisor | Reads each request and routes it; answers simple questions itself | Descriptions of all specialists, general FAQ |
| Billing Specialist | Invoices, refunds, subscription changes | Billing policy docs, invoice lookup tool |
| Technical Specialist | Troubleshooting and product questions | Product manuals, technical knowledge base |
| Order Status Specialist | Shipping and delivery questions | Order tracking tool via custom tools |
Example: phased outbound sales calls
Phase-based context control shines in structured voice conversations. An outbound calling workflow can run four phases, each handled by its own specialist that only sees what the phase requires:
1. Gatekeeper navigation
Reaches the decision-maker. Sees only the company and contact name — no product or pricing details, because they are not needed yet.
2. Problem discovery
Identifies the prospect's challenges. Cannot re-ask for basics collected in phase 1 and still has no access to pricing.
3. Solution presentation
Gets a summary of the discovered pain points plus relevant product information, and focuses only on matching a solution to them.
4. Objection handling and closing
Receives the full conversation summary and pricing, and works exclusively on next steps — it cannot loop back to earlier phases.
The result is a conversation that never repeats itself: each agent stays on topic, and the workflow progresses naturally from introduction to close.
Setting up a supervisor team
- 1
Create the specialists first
Build each sub-agent as a normal assistant with a narrow job, its own instructions, and only the knowledge and tools it needs.
- 2
Create the supervisor
Add a supervisor agent and attach the sub-agents to it. Give each sub-agent a clear one-line description — the supervisor routes based on these.
- 3
Define phases and context rules
For staged workflows, define the workflow phases and what information each agent may access in each phase.
- 4
Test the routing
In the playground, send requests that should reach each specialist and check that the supervisor routes them correctly and never re-asks for information it already has.
Best practices
- Keep specialists narrow: one clear job per sub-agent. Broad specialists defeat the purpose of routing.
- Write routing-friendly descriptions: the supervisor picks agents based on their descriptions, so state precisely what each one handles — and what it does not.
- Limit team size: start with two or three specialists. More agents mean more routing decisions and harder debugging.
- Scope knowledge and tools per agent: give each specialist only the files and tools it needs; shared everything makes answers less predictable.
- Follow the same instruction principles as single assistants — see the agent instructions guide.
Want help designing a multi-agent workflow for your team? Talk it through with an engineer.
Book a Call