What does it take to build a customer support experience your users won't hate? Ask Bitovi.

What this post added

This post details how Bitovi built a stateful AI agent for customer support using Temporal Workflows and Activities. The ReAct agent loop is implemented as a Temporal Workflow, with LLM inference and memory extraction handled by Activities. The post emphasizes the benefits of Temporal's Durable Execution for handling long-running, stateful conversations and integrating external memory systems like AWS Bedrock AgentCore Memory. It highlights the isolation of nondeterministic LLM inference within Activities and the use of structured output with validation and retries for reliability. The post also provides a code snippet for a `ThoughtActivity` and links to the full implementation on GitHub.