Subsystem 01
Omnichannel gateway
Ingests messages from Zendesk, Intercom, email and Slack
Typical stack
FastAPI + webhook handlers
Reference architecture
Autonomous customer support AI agent with tool execution
Stateful AI agents that resolve routine support tickets by querying your CRM and order systems, act only through approved tools, and hand off to a person when rules or confidence require it.
Design constraints
What the design has to hold to
Targets for the scenario this reference is sized for. A real engagement starts by replacing them with your own numbers.
- 01No account change or refund without the authorization that action requires
- 02Deterministic structured output validated against a strict JSON schema
- 03Escalation to a human agent with a full context summary
- 04First chat response within 2 seconds
Component topology
System components and technologies
Subsystems with separate responsibilities, clear contracts between them and storage that scales on its own. The stack named for each is typical, not mandatory.
Stack topology
Autonomous customer support AI agent with tool execution
Illustrative reference architecture
- 01
Omnichannel gateway
Ingests messages from Zendesk, Intercom, email and Slack
FastAPI + webhook handlers
- 02
Agent state orchestrator
Multi-step reasoning, memory and tool calling
LangGraph + Python
- 03
Knowledge & tool layer
Internal APIs for CRM lookups, order tracking and refunds
Internal tool registry
- 04
Human-in-the-loop queue
Interface where support staff approve high-value actions
Next.js dashboard + WebSockets
Subsystem 02
Agent state orchestrator
Multi-step reasoning, memory and tool calling
Typical stack
LangGraph + Python
Subsystem 03
Knowledge & tool layer
Internal APIs for CRM lookups, order tracking and refunds
Typical stack
Internal tool registry
Subsystem 04
Human-in-the-loop queue
Interface where support staff approve high-value actions
Typical stack
Next.js dashboard + WebSockets
Data lifecycle
End-to-end data flow
A customer asks a question in Intercom; the message is normalized and passed to the LangGraph agent.
The agent pulls the customer's history from Stripe and Zendesk through authenticated tool calls.
It plans a resolution; a refund above a set threshold (say $100) becomes a pending action in the human queue.
A support rep approves it in the Next.js dashboard, and the agent resumes and issues the refund through the Stripe API.
The agent drafts a confirmation with tracking links and closes the ticket.
Reliability and resilience
Failure modes and how each is contained
Failure mode 01
Prompt injection in customer messages
Mitigation
Treat customer text as untrusted data: keep it out of system instructions, give tools least-privilege scopes, and require approval for any action that moves money.
Failure mode 02
Agent decision loops
Mitigation
Cap tool calls per turn (for example 5) and fall back to human triage when the cap is hit.
Failure mode 03
Invented return-policy rules
Mitigation
Answer policy questions only from retrieved policy documents, with a link to the exact section.
Questions
What teams ask about this design
It depends on your ticket mix. Transactional requests such as order status, returns, password resets and billing updates are the usual candidates. We measure the share on a labeled sample of your own tickets before setting a target.
Through sentiment thresholds, minimum confidence scores and explicit business rules; for example, account cancellation requests always go to the retention team.
Planning a system like this?
Send us your requirements, expected load and budget. We'll reply within one business day with an honest read on the design, and on whether we're the right team to build it.