Agents dynamically transfer control to each other based on context.
Overview
The adaptive network pattern enables decentralized orchestration where agents autonomously decide when to hand off tasks to other specialists. Unlike the supervisor pattern, there’s no central coordinator—agents evaluate tasks and route them directly to peers.
User Request → Agent A → Agent B → Agent C → Response
└──────────┴──────────┘
Dynamic handoffs
When to Use
The adaptive network pattern is ideal when:
- Tasks naturally flow between domains.
- You need dynamic routing based on conversation context.
- Agents should autonomously decide when to hand off.
- Sequential expertise is required.
- You want lower latency than supervisor aggregation.
Good Fit Examples
| Use Case | Why Adaptive Network Works |
|---|
| Employee onboarding | HR → IT → Finance flow based on current step |
| Customer journey | Sales → Support → Success natural progression |
| Incident handling | Triage → Specialist → Resolution escalation |
| Complex troubleshooting | Problem domains emerge during conversation |
Architecture
┌─────────────────────────────────────────────────────────────────┐
│ Agentic App │
├─────────────────────────────────────────────────────────────────┤
│ │
│ ┌───────────────────────────────────────────────────────────┐ │
│ │ User Request │ │
│ └────────────────────────────┬──────────────────────────────┘ │
│ │ │
│ ▼ │
│ ┌───────────────────────────────────────────────────────────┐ │
│ │ Initial Agent │ │
│ │ (evaluates and may handle or hand off) │ │
│ └────────────────────────────┬──────────────────────────────┘ │
│ │ │
│ Can I handle this? ─────┐ │
│ │ │ │
│ ┌───────┴───────┐ │ │
│ ▼ ▼ │ │
│ ┌────────┐ ┌────────┐ │ │
│ │ Yes │ │ No │ │ │
│ └────┬───┘ └────┬───┘ │ │
│ │ │ │ │
│ ▼ ▼ │ │
│ ┌──────────┐ ┌───────────────────┐ │
│ │ Process │ │ Hand off to best │ │
│ │ Request │ │ suited agent │ │
│ └──────────┘ └─────────┬─────────┘ │
│ │ │
│ ┌───────────┼───────────┐ │
│ ▼ ▼ ▼ │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ Agent B │ │ Agent C │ │ Agent D │ │
│ └────┬─────┘ └────┬─────┘ └────┬─────┘ │
│ │ │ │ │
│ └────────────┴────────────┘ │
│ │ │
│ ▼ │
│ ┌───────────────────────────────────────────────────────────┐ │
│ │ Response │ │
│ └───────────────────────────────────────────────────────────┘ │
│ │
└─────────────────────────────────────────────────────────────────┘
Key Characteristics
Dynamic Delegation
Agents assess incoming tasks and route them to the most suitable peer when the task falls outside their scope.
User: "I can't log into my email"
IT Agent evaluates:
├── Is this an account issue? → Hand off to Identity Agent
├── Is this a device issue? → Handle directly
└── Is this a policy issue? → Hand off to HR Agent
Context Preservation
Conversation history travels with handoffs—users don’t repeat information.
Turn 1 (HR Agent): "I need to onboard a new employee, John Smith"
Turn 2 (IT Agent receives context): Already knows it's about John Smith
Turn 3 (Finance Agent receives context): Already knows full onboarding context
Decentralized Control
No central supervisor. Each agent is aware of other agents and can directly initiate transfers.
Sequential Workflows
Tasks typically flow through agents sequentially rather than in parallel.
Execution Flow
1. User submits initial query
2. Initial agent evaluates:
├── Intent analysis
├── Scope check
└── Capability assessment
3. Agent decision:
├── Handle directly → Process and respond
└── Hand off → Select best agent, transfer control
4. Next agent receives:
├── Original request
├── Conversation history
└── Handoff context
5. Process continues until resolution
6. Final agent delivers response to user
Supervisor vs. Adaptive Network
| Aspect | Supervisor | Adaptive Network |
|---|
| Control | Centralized | Decentralized |
| Execution | Parallel | Sequential |
| Routing | Supervisor decides | Agents decide |
| Latency | Higher (aggregation) | Lower (direct) |
| Complexity | Single coordination point | Distributed logic |
| Best for | Decomposable tasks | Flowing conversations |
Configuration
# app-config.yaml
orchestration:
pattern: adaptive_network
initial_agent: welcome_agent
# Handoff settings
handoff:
preserve_context: true
max_handoffs: 5
allow_circular: false
agents:
welcome_agent:
name: Welcome Agent
description: Greets users and routes to appropriate specialist
can_handoff_to:
- hr_agent
- it_agent
- finance_agent
hr_agent:
name: HR Agent
description: Handles HR inquiries, policies, and employee data
can_handoff_to:
- it_agent
- finance_agent
- welcome_agent
it_agent:
name: IT Agent
description: Manages IT support, access, and technical issues
can_handoff_to:
- hr_agent
- finance_agent
finance_agent:
name: Finance Agent
description: Handles payroll, expenses, and financial queries
can_handoff_to:
- hr_agent
Example: Employee Buddy
A “Smart Employee Buddy” where agents handle different domains and hand off as needed.
name: Smart Employee Buddy
orchestration:
pattern: adaptive_network
initial_agent: welcome_agent
agents:
- name: Welcome Agent
description: |
First point of contact. Greets employees and identifies
their needs. Routes to HR, IT, or Finance based on query.
instructions: |
You are the first contact for employee inquiries.
Quickly identify the domain and hand off to the right specialist.
Hand off rules:
- Benefits, leave, policies → HR Agent
- Computer, access, software → IT Agent
- Payroll, expenses, reimbursements → Finance Agent
- name: HR Agent
description: |
Handles leave requests, benefits inquiries, policy questions,
and employee data updates.
tools:
- get_leave_balance
- submit_leave_request
- get_benefits_info
handoff_triggers:
- "If user mentions IT issues, hand off to IT Agent"
- "If user asks about payroll, hand off to Finance Agent"
- name: IT Agent
description: |
Manages access requests, password resets, software installation,
and technical troubleshooting.
tools:
- reset_password
- request_software
- create_support_ticket
handoff_triggers:
- "If issue relates to HR systems, hand off to HR Agent"
- "If user needs expense report help, hand off to Finance Agent"
- name: Finance Agent
description: |
Handles payroll questions, expense submissions, reimbursements,
and tax documents.
tools:
- get_payslip
- submit_expense
- get_tax_documents
handoff_triggers:
- "If user has benefits questions, hand off to HR Agent"
Conversation Example
User: "Hi, I need some help"
Welcome Agent: "Hello! I'm here to help. What do you need
assistance with today?"
User: "I want to take vacation next week but I'm not sure
how much time I have left"
Welcome Agent: [Identifies HR domain, hands off to HR Agent]
HR Agent: "I can help with that! Let me check your leave balance."
[Calls get_leave_balance]
"You have 12 vacation days remaining. Would you like
me to submit a leave request for next week?"
User: "Yes please. Also, I realized I need to expense my
travel from last month"
HR Agent: [Submits leave request, then hands off to Finance Agent]
"Your leave request for next week is submitted.
I'm connecting you with our Finance team for your
expense report."
Finance Agent: "Hi! I see you need to submit a travel expense.
I can help with that. What was the trip for
and what's the total amount?"
Benefits
| Benefit | Description |
|---|
| Low latency | Direct agent-to-agent communication |
| Natural flow | Conversations progress organically |
| Agent autonomy | Specialists decide when to engage |
| Context continuity | Full history travels with handoffs |
| Scalability | Add new agents without central changes |
Best Practices
Clear Handoff Rules
Define explicit conditions for when agents should hand off:
handoff_triggers:
- condition: "User mentions billing or payment"
target: billing_agent
- condition: "Technical issue beyond my scope"
target: specialist_agent
- condition: "User explicitly requests different help"
target: router_agent
Prevent Circular Handoffs
Set maximum handoff limits and avoid circular routing:
handoff:
max_handoffs: 5
allow_circular: false
track_visited: true
Graceful Fallbacks
Handle cases where no agent can help:
fallback:
agent: support_agent
message: "I'll connect you with a specialist who can help."
