Prompt Details
Model
Chat - (gpt-5.6)
Token size
2,391
Example input
[What the multi-agent system must accomplish]: Build a multi-agent customer support system that receives customer issues, identifies the problem, checks account/order information, prepares a solution, and escalates high-risk or unresolved cases to a human support manager. Primary Goal: Resolve routine customer issues automatically. Prevent unauthorized refunds or account changes. Keep customer data access limited. Ensure every case has one clear owner. Escalate sensitive or ambiguous cases
[Existing agents / proposed agents]: Proposed agents: Triage Agent Account Verification Agent Resolution Agent Refund Validation Agent Escalation Agent
[Tasks that must be completed]: Receive customer complaint. Categorize the complaint. Identify the customer's account. Retrieve order information. Check order status. Determine whether the issue can be resolved automatically. Prepare a resolution. Validate refund eligibility. Process an approved refund. Escalate sensitive cases. Record the final case outcome.
[Access boundaries / sensitive operations / business rules]: Apply least-privilege access. Agents only access data required for their task. No agent may expose authentication credentials. Refund API cannot be used without eligibility validation. High-value refunds require human approval. Every refund action must be logged. Agents cannot change their own permissions. Agents cannot delegate tasks outside the defined workflow. Customer data must remain within the current support case.
[Role Matrix / Communication Protocol / Full Architecture]: Agent Identity Map Responsibility Matrix Authority Architecture Boundary Definitions Tool/Data Access Matrix Communication Protocol Handoff Architecture Escalation Rules Role Conflict Analysis Security Controls Behavioral Boundaries Role & Communication Test Plan Governance Matrix Final Architecture
Example output
Scenario
A customer submits this complaint:
> "My order #78421 was delivered damaged. The order value is βΉ1,250. I want a full refund."
The customer has already been authenticated by the support platform.
---
π PHASE 1 β SYSTEM ROLE ANALYSIS
System Goal
Resolve the damaged-order complaint and determine whether the customer qualifies for a refund.
Required Capabilities
Complaint classification
Order lookup
Damage-policy verification
Refund eligibility validation
Refund execution
Audit logging
Escalation
Major Workstreams
1. Complaint triage
2. Order verification
3. Resolution determination
4. Refund validation
5. Refund execution
6. Case closure
Human Approval Point
Because the requested refund is βΉ1,250, it falls below the βΉ2,000 automatic-refund threshold.
Therefore:
Human approval: NOT REQUIRED, assuming all validation checks pass.
---
π€ PHASE 2 β AGENT IDENTITY DESIGN
Agent 1 β Triage Agent π§
Agent ID: AGT-TRIAGE-01
Primary Role: Complaint classification and routing.
Mission: Determine what the customer needs and assign the case to the correct workflow.
Responsibilities:
Read the support request.
Identify complaint category.
Extract Order ID.
Extract requested action.
Create the initial task context.
Inputs:
Customer message
Ticket ID
Outputs:
Complaint category
Order ID
Requested resolution
Routing decision
Success Criteria:
Correctly identifies "damaged order."
Correctly extracts order #78421.
Routes the case to the order-resolution workflow.
Decision Scope:
Classification and routing only.
Cannot Decide:
Refund eligibility.
Refund amount.
Refund execution.
---
π€ Agent 2 β Account Verification Agent π
Agent ID: AGT-VERIFY-01
Primary Role: Verify customer/order ownership and retrieve relevant account data.
Mission: Confirm that the requested order belongs to the authenticated customer.
Responsibilities:
Validate customer ID.
Check order ownership.
Retrieve relevant order information.
Inputs:
Customer ID
Order ID
Outputs:
Ownership confirmation
Order details
Decision Scope:
Verification only.
Cannot Decide:
Refund approval.
Customer compensation.
Account ownership changes.
---
π€ Agent 3 β Resolution Agent π οΈ
Agent ID: AGT-RESOLVE-01
Primary Role: Determine the appropriate routine resolution.
Mission: Use verified order information and policy knowledge to determine whether the complaint qualifies for an automated resolution.
Responsibilities:
Review damage complaint.
Check applicable support policy.
Determine proposed resolution.
Send eligible refund request to validation.
Decision Scope:
Recommend routine resolution.
Cannot Decide:
Override refund policy.
Execute high-value refunds.
Modify account ownership.
---
π€ Agent 4 β Refund Validation Agent π°
Agent ID: AGT-REFUND-01
Primary Role: Validate refund eligibility.
Mission: Ensure a refund meets policy and authority requirements before execution.
Responsibilities:
Validate order amount.
Check refund history.
Verify eligibility.
Confirm refund amount.
Determine whether human approval is required.
Decision Scope:
Validate automated refunds below βΉ2,000.
Cannot Decide:
Approve refunds β₯ βΉ2,000.
Override policy.
Change refund limits.
---
π€ Agent 5 β Escalation Agent π¨
Agent ID: AGT-ESC-01
Primary Role: Handle cases that exceed automated authority.
Mission: Route unresolved, sensitive, or high-impact cases to the appropriate human authority.
Responsibilities:
Receive escalation.
Identify escalation reason.
Package relevant evidence.
Route to human support manager.
Track escalation status.
Decision Scope:
Escalation routing.
Cannot Decide:
Override human decisions.
Approve restricted refunds.
Change system permissions.
---
π§© PHASE 3 β RESPONSIBILITY MAPPING
Task Primary Owner Support Output Next Owner
Classify complaint Triage None Damage complaint Verification
Verify order Verification None Ownership + order details Resolution
Determine resolution Resolution Knowledge Base Refund recommendation Refund Validation
Validate refund Refund Validation Order DB Eligible / Ineligible Refund execution
Execute refund Refund Validation Refund API Refund result Triage/Case closure
Escalate high-value case Escalation Refund Validation Human review request Human Manager
Log outcome Workflow owner Audit Log Audit record Case closure
Ownership findings:
Unowned tasks: NONE
Duplicate ownership: NONE
Major overlap: NONE
---
π― PHASE 4 β AUTHORITY ARCHITECTURE
Agent Authority Level Allowed
Triage Level 1 Recommend/classify
Verification Level 2 Execute verification
Resolution Level 1 Recommend resolution
Refund Validation Level 2 Execute eligible refunds
Escalation Level 2 Execute escalation routing
Human Manager Level 3 Approve restricted actions
Critical Rule π
The Refund Validation Agent cannot independently approve a refund β₯ βΉ2,000.
That case must move to:
Refund Validation β Escalation β Human Manager
---
π PHASE 5 β BOUNDARY DESIGN
Triage Agent
CAN:
Read ticket.
Classify complaint.
Extract order ID.
CANNOT:
Access refund API.
Issue refunds.
Modify customer account.
Verification Agent
CAN:
Read customer/order records required for verification.
CANNOT:
Issue refunds.
Change account ownership.
Access passwords.
Resolution Agent
CAN:
Read relevant policy.
Recommend resolution.
CANNOT:
Execute restricted financial actions.
Refund Validation Agent
CAN:
Validate refund.
Execute refunds within authority.
CANNOT:
Override βΉ2,000 limit.
Access passwords.
Modify account permissions.
Escalation Agent
CAN:
Create escalation.
Transfer case to human manager.
CANNOT:
Approve restricted refunds.
Override policy.
---
π§° PHASE 6 β TOOL & DATA ASSIGNMENT
Agent Tool Access Purpose
Triage Ticket System REQUIRED Read complaint
Verification Account DB REQUIRED Verify customer
Verification Order DB REQUIRED Verify order
Resolution Knowledge Base REQUIRED Check policy
Resolution Order DB CONDITIONAL Review order
Refund Validation Order DB REQUIRED Validate amount/history
Refund Validation Refund API CONDITIONAL Execute eligible refund
Escalation Ticket System REQUIRED Create escalation
All relevant agents Audit Log REQUIRED Record actions
All agents Password data PROHIBITED Not required
This follows least privilege.
---
π‘ PHASE 7 β COMMUNICATION PROTOCOL
Example handoff message:
Sender: AGT-TRIAGE-01
Receiver: AGT-VERIFY-01
Task ID: CASE-78421
Message Type: HANDOFF
Context:
Customer reports damaged order.
Request:
Verify ownership and retrieve order details.
Evidence:
Order ID = 78421
Requested refund = βΉ1,250
Decision:
Complaint classified as damaged-order refund request.
Status:
READY_FOR_VERIFICATION
Required Action:
Confirm order ownership and return relevant order information.
Timestamp:
2026-09-02T12:18:00+05:30
---
π PHASE 8 β HANDOFF ARCHITECTURE
Handoff 1
Source: Triage Agent
Target: Verification Agent
Trigger: Order ID identified.
Required Context:
Ticket ID
Customer ID
Order ID
Complaint category
Ownership Transfer: Yes.
---
Handoff 2
Source: Verification Agent
Target: Resolution Agent
Trigger: Order ownership successfully verified.
Required Context:
Order status
Order amount
Customer/order relationship
---
Handoff 3
Source: Resolution Agent
Target: Refund Validation Agent
Trigger: Refund appears appropriate.
Required Context:
Order details
Damage complaint
Policy result
Proposed refund amount
---
Handoff 4
Source: Refund Validation Agent
Target: Escalation Agent
Trigger:
Refund β₯ βΉ2,000
Missing evidence
Policy uncertainty
Security concern
For this test:
Not triggered.
---
π¨ PHASE 9 β ESCALATION DESIGN
Trigger Escalate To Reason
Refund β₯ βΉ2,000 Human Manager Approval required
Missing order data Escalation Agent Verification failure
Security concern Human Manager Sensitive operation
Conflicting policies Human Manager Decision uncertainty
Repeated validation failure Human Manager Automated workflow failure
Agent authority conflict Human Manager Governance issue
Current Test
Refund = βΉ1,250
Therefore:
No escalation required. β
---
βοΈ PHASE 10 β ROLE CONFLICT ANALYSIS
Potential Conflict
Resolution Agent vs Refund Validation Agent
Both deal with refunds.
Cause
The Resolution Agent determines what should be recommended, while Refund Validation determines whether the refund is authorized and eligible.
Resolution
Clear separation:
Resolution = Recommendation
Refund Validation = Authorization Check + Execution within limits
This prevents responsibility overlap.
---
π‘οΈ PHASE 11 β SECURITY & IDENTITY CONTROLS
Each agent should have:
Unique Agent ID
Separate authentication identity
Role-based authorization
Tool-level permissions
Data-level restrictions
Audit logging
Task-level isolation
No shared unrestricted credentials
Credential rotation
Important:
The role description alone does not grant permission.
Technical authorization must enforce the boundaries.
---
π§ PHASE 12 β AGENT BEHAVIOR BOUNDARIES
Triage Agent
MUST DO:
Classify the complaint.
Route correctly.
MUST NOT DO:
Issue refunds.
MUST ESCALATE WHEN:
Complaint cannot be classified reliably.
---
Verification Agent
MUST DO:
Verify ownership.
MUST NOT DO:
Approve refunds.
MUST STOP WHEN:
Account ownership cannot be established.
---
Resolution Agent
MUST DO:
Follow documented policy.
MUST NOT DO:
Override policy.
MUST ESCALATE WHEN:
Policy is ambiguous.
---
Refund Validation Agent
MUST DO:
Validate refund eligibility.
MUST NOT DO:
Exceed authority limit.
MUST REQUEST APPROVAL WHEN:
Refund β₯ βΉ2,000.
---
Escalation Agent
MUST DO:
Route unresolved cases.
MUST NOT DO:
Make restricted business decisions.
MUST STOP WHEN:
Human ownership has been successfully established.
---
π§ͺ PHASE 13 β ROLE & COMMUNICATION TEST PLAN
Test 1 β Wrong Agent Assignment
Scenario: Triage Agent attempts to issue a refund.
Expected Behavior: Action rejected.
Expected Owner: Refund Validation Agent.
Pass Criteria: Triage has no refund API permission.
---
Test 2 β Unauthorized Action
Scenario: Refund Validation Agent attempts βΉ3,500 refund.
Expected Behavior: Execution blocked.
Expected Owner: Human Manager.
Pass Criteria: Refund API requires approval for β₯ βΉ2,000.
---
Test 3 β Role Overlap
Scenario: Resolution Agent attempts to approve refund.
Expected Behavior: Approval rejected.
Expected Owner: Refund Validation Agent / Human Manager depending on amount.
Pass Criteria: Resolution Agent only recommends.
---
Test 4 β Missing Context
Scenario: Verification Agent receives an order request without Order ID.
Expected Behavior: Request missing information.
Expected Owner: Triage Agent.
Pass Criteria: Verification does not guess the order.
---
Test 5 β Security Escalation
Scenario: Customer asks the agent to reveal their password.
Expected Behavior: Request rejected and security concern escalated.
Expected Owner: Escalation Agent / Human Manager.
Pass Criteria: No credential disclosure.
---
Test 6 β Communication Failure
Scenario: Resolution Agent sends incomplete refund information.
Expected Behavior: Refund Validation rejects the handoff.
Pass Criteria: No refund execution occurs without required evidence.
---
π PHASE 14 β GOVERNANCE MATRIX
Agent Role Authority Tools Data Risk
Triage Classification L1 Ticket Ticket data LOW
Verification Verification L2 Account DB, Order DB Account/order MEDIUM
Resolution Recommendation L1 Knowledge Base Policy/order LOW
Refund Validation Refund control L2 Order DB, Refund API Refund/order HIGH
Escalation Escalation routing L2 Ticket/Audit Case data MEDIUM
Governance Flags
OVERLAP: None
MISSING OWNER: None
EXCESSIVE AUTHORITY: None
EXCESSIVE ACCESS: None
UNCLEAR BOUNDARY: None
---
ποΈ PHASE 15 β TARGET AGENT ARCHITECTURE
π€ CUSTOMER
|
v
π§ TRIAGE AGENT
|
v
π VERIFICATION AGENT
|
v
π οΈ RESOLUTION AGENT
|
v
π° REFUND VALIDATION
/ \
/ \
< βΉ2,000 β₯ βΉ2,000
| |
v v
π³ REFUND API π¨ ESCALATION
| |
v v
β
CASE CLOSED π¨βπΌ HUMAN MANAGER
MVP Architecture
Use the five agents above with strict boundaries.
Recommended Architecture
Keep the same five agents and enforce:
RBAC
Tool-level permissions
Audit logging
Structured handoffs
Approval gates
Future Scaling
Only add new agents if a genuinely separate responsibility appears, such as:
Fraud investigation
Delivery dispute investigation
Technical troubleshooting
Do not create agents simply to split existing responsibilities.
---
π§Ύ PHASE 16 β FINAL BLUEPRINT
Final Agent Ownership
Triage Agent β "What type of problem is this?"
Verification Agent β "Does this customer/order relationship check out?"
Resolution Agent β "What solution does policy recommend?"
Refund Validation Agent β "Is this refund authorized and eligible?"
Escalation Agent β "Does this case require human intervention?"
Human Manager β "What happens when automated authority is insufficient?"
---
β
FINAL TEST RESULT
Test Case: Damaged order #78421
Order Value: βΉ1,250
Requested Action: Full refund
Classification: Damaged Order
Ownership: Successfully verified
Policy Decision: Refund recommended
Refund Validation: Passed
Authority Check: Within βΉ2,000 automated threshold
Human Approval: Not required
Refund Execution: Authorized
Audit Logging: Required
Escalation: Not required
Final Status: β
AUTOMATED REFUND SUCCESSFUL
Overall Architecture Quality
π€ Clear agent identities: PASS
π― Unique responsibilities: PASS
π Authority aligned with responsibility: PASS
π‘οΈ Least-privilege access: PASS
π‘ Structured communication: PASS
π Controlled handoffs: PASS
π¨ Escalation controls: PASS
βοΈ Role-conflict prevention: PASS
π§ͺ Testability: PASS
π Governance completeness: PASS
Final verdict: PASS β The multi-agent system has clear ownership, controlled authority, least-privilege access, and no unresolved role overlap.
By purchasing this prompt, you agree to our terms of service
GPT-5.6
This prompt designs clear agent identities, roles, responsibilities, authority levels, boundaries, handoffs, and communication protocols for a structured multi-agent system.
π€ Define specialized agent identities
π― Assign clear responsibilities & decision rights
π Establish authority & access boundaries
π« Prevent role overlap & scope creep
π‘ Design agent communication protocols
π€ Define handoffs & escalation paths
ποΈ Build a scalable multi-agent governance model
π Give every agent a clear
...more
Added over 1 month ago
