본문 바로가기
Opinion · 2026.10.08

Let the LLM only decide what to do next and hand execution to a deterministic harness — cluster restarts then run the same way every time

Brains vs Hands: How to Run AI Agents Safely in Production — Viren Baraiya — AI Engineer · Read the original ↗

This page is a translation of the Korean original. Read the Korean original →

What you get

  • Risky operations such as production restarts and rollbacks follow the same procedure on every run
  • Human approval gates are enforced by the harness's guaranteed execution, not by the model's judgment
  • Loops that run for days or months recover after failures, and their side effects are recorded

When to use it

Apply it when agents touch operational work that must run identically every time, such as restarting production clusters or rolling back deployments. It also fits background agents that react to log or alert events or run on a schedule, and long-running harnesses that wait days or months for responses from people or external services.

How to do it

  1. Design the agent not as a single component but as an application (a harness) in which systems, people and tools run together.
  2. Inside the harness, clearly separate the non-deterministic part (LLM reasoning) from the deterministic part (operations such as payments or cluster restarts).
  3. On every loop, maintain a 'state of the world' that holds completed work, recorded side effects and success or failure, and have the LLM plan the next step from that state and the goal.
  4. For risky operations such as restarting a production cluster, add a human approval gate, such as a Slack message, that the harness is guaranteed to execute.
  5. Make steps idempotent; where a step cannot be idempotent, record what happened and handle the side effects separately.

Where I agree or hold back

The case for drawing a line between planning and execution is worth adopting as is, especially the point that approval gates and idempotency should not be left to the model's judgment. The premise that 'you cannot script every branch in advance, so the agent assembles them at runtime' is one to hold back on, though: for tasks with few tools and predictable paths it may be more structure than you need. The speaker also acknowledged that the demo was pre-scripted, so measure it yourself.

Background — what the original covers

In this AI Engineer World's Fair 2026 talk, Viren Baraiya, who built Netflix Conductor, discusses running agents in production. Once you look past chatbot examples to scheduled, event-driven, long-running and multi-agent setups, an agent looks less like a component and more like an application, and the harness around it starts to resemble a bundle of microservices. He calls this a 'late-bound saga' assembled at runtime and demonstrates it with an SRE agent.

#Agents#Safety

← All practitioner tips