Living in Color W33: What Is a Harness?

Beginner week. No shame in the question. Half the timeline uses “harness” like everyone already agreed on a definition. They did not.

Plain language

A model is the brain you rent or host.

A harness is the workbench around that brain: the CLI or IDE, the tool permissions, the memory, the loop that lets it read files, run commands, open PRs, and stop when you say stop.

An agent is what you get when a model is wired through a harness with a job and tools. The agent is not the logo. It is the whole system.

If you only buy “the smartest model” and paste into a chat box, you bought a brain with no hands.

Model vs harness vs agent (one picture in words)

Think of a kitchen.

• Model = the cook’s skill
• Harness = the kitchen: knives, stove, recipes on the wall, who has the key
• Agent = tonight’s service: skill + kitchen + tonight’s tickets

Claude Opus inside Claude Code is a different agent than Claude Opus inside a raw API script, even when the model string matches. Same cook, different kitchen.

That is why the Bingo Board grades models and harnesses as separate layers. A mid model in a great harness can beat a frontier model in a toy chat UI for real shipping work.

The current set I point beginners at: Claude Code, Codex, Cursor, Antigravity, Grok Build, and Muse Code. Pick one. Do not collect all six.

How the Board grades harnesses

I look at:

1. Tool surface - files, shell, browser, MCP, PR flow
2. Control - approvals, resumable runs, audit of what changed
3. Plan economics - can I live here daily without surprise bills
4. Model access - which brains can I actually route
5. Feel under load - does it survive a multi-hour engineering day

Claude Code is the most complete control surface I have used (Skills, Subagents, Hooks, MCP, Plan). I do not open it first every morning. Cursor Ultra is the daily seat. Codex is knowledge work, GTM, and the OpenAI coding ceiling in a terminal. OpenCode takes Linear 1-3. Antigravity is where Google’s Flash computer-use story earns rent. Grok Build is architecturally interesting and still a hard sell on price vs receipts. Muse Code (Meta, Aug 5) is the new watch harness: persistent background agents, worktree fan-out, an append-only event log, and skills. No Chris grade yet.

Minimum viable harness if you are starting this month

Do not buy five subscriptions.

1. Pick one daily harness you will open every morning.
2. Write a one-page tool allowlist (what it may touch in prod vs staging).
3. Pick a default model and a ceiling model (cheap daily vs hard ticket).
4. Add a stop rule: human approval before write access to prod or customer data.
5. Review the allowlist every Monday for ten minutes.

That is enough to be a builder. Everything else is optimization.

Short signal coda

• Board since mid-month: no letter moved. The four August rows are still ungraded watch. The live standouts now say what I run (Cursor Ultra + Codex), not Opus 4.8 as the daily.
• Harness update: Muse Code is still the control-surface story (event log, worktree fan-out, muse resume). I have not made it a daily.
• Ignore: the Aug 6 pulse still listing GPT-6 and Composer 3 like they are on a calendar. That is chatter, not a Board event.

One builder action

Open your current harness and list every MCP server or tool plugin enabled. Delete one you cannot explain to a security person in one sentence.

Companion vlog: What is a coding harness? Plain language.