The State Machine Behind the Agent | Making LLM Workflows Reliable

Speaker: Johnkingsley Amaechi AI agents are everywhere and are easier than ever to create. The hard problem with agent use is controlling what happens after they start executing: calling tools, waiting on external systems, retrying failed steps, carrying context and deciding when a human should approve the next action. This talk discusses how the speaker used state machines as a control layer behind the agent to model workflow states, add checkpoints and place human approval gates where automation should stop. And for their use case, this made agent workflows easier to debug, recover, audit, and trust.