Let Agents Run Wild — Inside Deterministic Guardrails
AI agents are non-deterministic. They hallucinate, vary every run, and surprise you. The industry keeps trying to constrain them. I take the opposite approach: let agents run fully autonomous inside deterministic durable function workflows that guarantee the pipeline always completes. You already know how to operate deterministic workflows. Now use that skill to tame AI.
AI agents in production will hallucinate, take paths you didn't plan for, timeout, contradict themselves, and occasionally return poetry when you asked for JSON. And that's fine. As long as your orchestration is deterministic.
We stopped fighting non-determinism and started embracing it. We took our AI agents off the leash and wrapped them in durable function workflows that provide the guardrails: checkpoints, retries, timeouts, exactly-once execution. The agent can go wild inside each step. The orchestrator guarantees the pipeline completes.
I walk through a real system where multiple AI agents reason and act autonomously. They take unexpected paths. They fail in creative ways. But the durable function workflow keeps the train on the tracks. It checkpoints state, fans out for parallel work, reconciles results, and recovers from failures automatically.
AI in production is not a new operational paradigm. It's a non-deterministic step inside a deterministic workflow. And you already know how to operate deterministic workflows.
Eric Johnson is a Principal Developer Advocate at AWS specializing in serverless and event-driven architectures. With 35+ years in technology and a passion for making the complex understandable, Eric has presented and keynoted at hundreds of tech conferences worldwide. He brings humor, energy, and hard-won production experience to every talk, and has been an early adopter of AWS durable functions for agent orchestration since their launch.