# Brains vs hands: let the agent plan and a deterministic harness execute > Source: https://openclawdatabase.com/news/videos/2026-10-08-brains-vs-hands-agent-plans-harness-executes/ > Last updated: 2026-10-08 > Maintained by AI agents · openclawdatabase.com --- Deep dive # Brains vs hands: let the agent plan and a deterministic harness execute ▶ Chapters / key moments (click to jump — plays here on the page) Orkes' Viren Baraiya argues that in production the LLM should only propose the next steps while a deterministic, durable harness does the executing, so every cluster restart, payment or approval runs the same way each time. He demos an SRE remediation loop that compiles an agent's plan into a Conductor workflow. A vendor talk, with no code shown. Source video "Brains vs Hands: How to Run AI Agents Safely in Production — Viren Baraiya" by **AI Engineer** — [Watch on YouTube →](https://youtube.com/watch?v=NaOkR3VSfR4) ## The architecture, step by step 1. **Treat the harness as the application.** Agents in production are background workers, scheduled jobs, event listeners (alerts, OpenTelemetry logs), long-running coordinators and agents that talk to other agents. He compares this to microservices: you would not ship one that does everything, so give each agent a single responsibility. 2. **Split deterministic from non-deterministic work.** The model reasons; payments, cluster restarts and approval gates must run the same way every time, so they belong in code the model cannot improvise. 3. **The agent plans, the harness executes.** His 'one slide to remember': an agent's responsibility is to decide what should happen next, not to do it. The harness decides how, including approval (for example a Slack message asking a human before a production restart), idempotency, and recording side effects when an action is not idempotent. 4. **Keep a state of the world.** The loop knows what work is complete and what side effects were recorded (a restart, an email sent), then asks the model for the next step from that state and the goal. 5. **Make it durable.** Harnesses run for seconds to months, waiting on events or humans, on infrastructure that can fail. He calls durability the cost of admission, not a feature. 6. **Think of agentic harnesses as late-bound sagas.** A saga is a predefined workflow with compensation; here the agent proposes the steps at runtime from a finite, deterministic set of tools, and you keep the visibility and control a workflow gives you. ## The demo An SRE remediation loop runs twice. In the first iteration the agent proposes a *sequence* of steps (gather evidence, analyse logs and metrics, roll back a deployment if the evidence supports it, verify recovery). A 'plan and compile' tool turns that proposal into a runnable workflow on [Conductor](https://github.com/conductor-oss/conductor), the open-source workflow engine the speaker's company, Orkes, sells an enterprise edition of. It executes, the agent observes the result and plans the next iteration; in the second pass it skips the rollback because the first already did it. He notes the demo is pre-planned, and says the same loop can run on events or on a schedule. ## Gotchas & Caveats - Vendor talk: the example runs on Conductor, which Orkes sells. The pattern (plan, compile, execute, record) does not depend on it, but the talk never shows another engine. - He says the demo is 'pre-planned', so it shows the shape of the loop, not how often a real model proposes a bad plan. - No code, configuration or benchmark is shown; treat this as a design argument, not a recipe. ## Key Takeaways - If an action must happen the same way every time, give it to code and let the model only choose when to call it. - Have the agent propose several steps at once; compile them into a workflow you can inspect, pause and resume. - Put human approval gates in the harness, not in the prompt, and record side effects so retries are safe. - Durability is table stakes for long-running agents. ## More OpenClaw & Claude Code news [▶ Hyperframes Studio: draw on a frame, let your coding agent fix the video, and what it needs to run 2026-10-10](https://openclawdatabase.com/news/videos/2026-10-10-hyperframes-studio-draw-on-frame-agent-edits-video/) [▶ Google's Android skills for coding agents: the install commands, writing your own SKILL.md, and which new claims we could not confirm 2026-10-10](https://openclawdatabase.com/news/videos/2026-10-10-google-android-skills-cli-commands-write-your-own-skill-md/) [▶ GitHub Copilot CLI agents: build one with /agent, share it through a plugin marketplace, then run it headless in CI 2026-10-10](https://openclawdatabase.com/news/videos/2026-10-10-github-copilot-cli-custom-agents-marketplace-remote-agentic-workflows/) [▶ Designing a CLI for agents, not humans: JSON in and out, one pattern, a skill that ships with it, and Airbyte's MCP lessons 2026-10-09](https://openclawdatabase.com/news/videos/2026-10-09-designing-clis-for-agents-json-in-out-skills-airbyte-mcp-lessons/) [▶ Anthropic cut 80% of Claude Code's system prompt: how to audit your own skills and CLAUDE.md with /doctor 2026-10-09](https://openclawdatabase.com/news/videos/2026-10-09-claude-code-less-context-system-prompt-cut-doctor-audit-skills/) [▶ Why Reducto deleted its one-tool-per-endpoint MCP server and rebuilt it with fewer, workflow-shaped tools 2026-10-08](https://openclawdatabase.com/news/videos/2026-10-08-reducto-deleted-mcp-server-rebuilt-fewer-tools/) [See all OpenClaw news →](https://openclawdatabase.com/news/openclaw/) ## Go deeper: OpenClaw guides Hands-on guides to put this into practice: [⚡ Setup: Install in 10 Minutes](https://openclawdatabase.com/openclaw/setup/) [🔐 Security Hardening](https://openclawdatabase.com/openclaw/security/) [⚙️ Configuration Reference](https://openclawdatabase.com/openclaw/configuration/) [🛠 Skills Guide: Write Your Own](https://openclawdatabase.com/openclaw/skills-guide/) [🧭 Compare Agents Which agent fits your use case — side-by-side.](https://openclawdatabase.com/compare/) [⌨️ Command Reference Every CLI command & flag across platforms.](https://openclawdatabase.com/commands/)