Customer support agents need to do more than generate text. They need to access customer data, follow business rules, remember conversations, and handle sensitive actions safely.
To explore this, I built Smart Support Agent using Strands Agents, Amazon Bedrock, and Amazon Bedrock AgentCore Runtime.
🛠️ What I Built
The agent can:
- Look up customer accounts
- Apply steering and safety rules
- Delegate technical issues to a specialist agent
- Evaluate responses and tool-call workflows
The main customer-service tools are:
lookup_customerget_order_historyprocess_refund
🔐 Making Refunds Safer
One of the most valuable parts of the project was designing the refund workflow.
The agent doesn't immediately process a sensitive request. It first verifies the customer and checks the relevant order.
During testing, a partial $9.99 refund was proposed for a $79.99 order without sufficient context. Instead of blindly executing it, the steering layer blocked the action and required human review.
I then tested the complete valid workflow:
Customer C-1001 → Order ORD-5521 → Full refund of $79.99
The deployed AgentCore runtime successfully processed the refund.
This demonstrated an important principle:
High-impact agent actions should have deterministic controls around them, not rely only on the LLM.
💾 Solving Session Persistence
During deployment, I encountered:
PermissionError: [Errno 13] Permission denied: '/var/sessions'
The application needed a writable location for file-based session storage.
I changed the session directory to:
/tmp/smart-support-sessions
and kept the conversation manager for maintaining the agent's context.
This was a useful reminder that local development environments and cloud runtimes can have different filesystem constraints.
🤖 Multi-Agent Technical Support
I then extended the system using an agents-as-tools approach.
The main customer-service agent can delegate technical problems to a dedicated Tech Support Specialist.
Customer ↓Customer Service Agent ├── Account / Orders → Customer Tools ├── Refund → Verified Refund Workflow └── Technical Issue ↓ Tech Support Specialist ↓ Diagnostics & Troubleshooting
For example, when a customer reported that their wireless headphones were not pairing, the orchestrator delegated the issue to the specialist.
The test produced:
[DELEGATION] Tech support specialist activated[DELEGATION] Specialist responded
🧪 Evaluating Agent Behavior
I also explored automated evaluation for the agent.
Output evaluation checks:
- Professional and empathetic tone
Trajectory evaluation checks whether the agent follows the expected tool workflow.
For a refund, the expected sequence is:
lookup_customer ↓get_order_history ↓process_refund
This matters because an agent can produce a convincing final answer while still following an unsafe workflow internally.
☁️ Deploying to Amazon Bedrock AgentCore
Runtime.
The runtime reached:
smartSupportAgent: Deployed - Runtime: READY
I then successfully invoked the deployed AWS runtime and tested customer lookup, order retrieval, and the complete refund workflow.
After local testing, I packaged and deployed the application to Amazon Bedrock AgentCore :-
💡 What I Learned
This project taught me that building an AI agent isn't just about getting an LLM to generate a good answer.
Reliability comes from the system around the model:
Tools + Guardrails + Persistence + Delegation + Evaluation + Cloud Runtime
The most valuable part was debugging real deployment issues and understanding how these components work together.
🚀 What's Next?
Some improvements I would like to explore next:
- A user-friendly web chat interface
- Human approval for high-value refunds
- Better observability and monitoring
🔗 Source Code
The complete project is available on GitHub:
Building Smart Support Agent gave me a much deeper understanding of how to move from a simple AI assistant to a tool-using, safety-aware, multi-agent application deployed on AWS.
Loading article