Kafka as the Agent Action Log: Replay, Recovery, and Observability for AI Agents

Agentic AI systems don’t just need prompts and models. They need a durable record of what happened: what the agent tried, which tools it called, what failed, what was retried, and what state changed along the way.

That starts to look a lot like a log.

In this talk, I’ll show how Kafka can be used as an Agent Action Log: a Kafka-native pattern for recording agent execution as a stream of structured events. Tool calls, model responses, state transitions, retries, failures, and outputs are written to topics so agent workflows become easier to observe, replay, debug, and recover.

I’ll walk through a working implementation, including how to model tool calls as events, how agent state can live in Kafka, what happens when an agent stalls mid-task, and why idempotency gets harder when your worker is non-deterministic. I’ll also cover the tradeoffs: where Kafka fits naturally, where it gets awkward, and where purpose-built workflow engines or tracing systems may still make more sense.

You’ll leave with a concrete architecture for using Kafka to make AI agents more observable and recoverable, plus a pattern you can sketch on a whiteboard and start experimenting with the same day.

By attending this session, your contact information may be shared with the sponsor for relevant follow up for this event only.