Brains vs. hands: how to build production-safe AI agents

The core claim. The speaker argues that production-ready AI agents require a strict separation between planning and execution. The plan is the non-deterministic output of an LLM. The execution is a deterministic, durable workflow that should never leave room for hallucination. He calls this the brains vs. hands distinction.

Why the analogy matters. The speaker draws a direct parallel to microservices architecture. Just as you would not deploy a single microservice that does everything, you should not deploy a single agent that reasons, plans, and acts autonomously. Instead, a harness (the hands) should own all execution, and the LLM (the brains) should only plan what needs to happen next.

The specific failure mode. Without this separation, a critical action like a production cluster restart could be made by an LLM without a human gate. The speaker gives the example of a harness that ensures every cluster restart follows the exact same deterministic steps. He argues that leaving approval gates to an LLM introduces unacceptable risk.

The operational requirement. The harness must be durable and long-running. It has to survive network failures, partition events, and infrastructure restarts. Durability, he says, is table stakes, not a feature. The harness must also be idempotent where possible, and when it is not, it must record side effects so they can be handled separately.

The technical implementation. The speaker demonstrates a Conductor-based harness that ingests an LLM's multi-step plan and compiles it into a deterministic workflow. The demo shows an SRE remediation loop that runs twice: first to diagnose and roll back, then to verify recovery. The harness decides not to roll back on the second iteration because it was already done.

What this means for teams. The talk is framed as an engineering strategy, not a product pitch, though Conductor is the underlying tool. The insight is that production agents should be treated as late-bound sagas: a finite set of deterministic tools that an LLM can compose at runtime. Teams looking to run agents in production should focus on the harness architecture first, not the model.