Agent Harness vs Coding Harness (Know the Difference)

Stop blaming the model — your harness is the bottleneck, and the loop you're in decides which tool wins.

Two loops, two tool classes

Coding harnesses (Claude Code, Kilo Code, Cursor) sit in the dev loop: repos, IDEs, terminals, git, CI. A coding agent's authority is bounded to files and test runners, and a human still owns the final merge and release path.

General agent harnesses (Hermes, OpenClaw) sit in the runtime loop: sessions, tools, memory, cron, channels, routing. Hermes describes itself as "not a coding co-pilot tethered to an IDE" — it is an autonomous agent on your server that gets more capable the longer it runs.

Deterministic vs. inferential

The technical split is execution style. Coding harnesses run on deterministic execution loops — they verify by actually running the code. Anthropic's own harness test uses Playwright MCP to click through full-stack apps, hitting UI features, API endpoints, and database states end-to-end. Agent harnesses are inferential ("if it looks right, then it's right") and persist guesses across sessions.

The Pine Script demo

Same strategy, same model (Kimi 2.6). On Hermes, the indicator drew a single line from the open of the hour — nothing actionable, multiple syntax errors, and far fewer variables in the script. On Kilo Code, the same prompt produced a structured strategy: brown half-hour windows for entry, a structure-break filter, and stops below the recent low. Pine Script is a specialized DSL, and the harness's toolchain decided the quality gap.

Pick by job, not by model

• Scripts, persistent workflows, research → Hermes
• Multi-channel automation → OpenClaw
• Full-stack apps with testable backends, starting from scratch, bring-your-own-key → Kilo Code
• Refactor of an existing working codebase → Claude Code

The cost nobody mentions

To get an agent harness coding as well as a coding harness on the same model, you need "hundreds or thousands of hours" of feedback if you have no engineering background. With engineering experience, you can cut that roughly in half by teaching it branch conventions directly. The channel is close to 10K subscribers.

Caveat

The creator flags this as a mental model, not a specification — real harness design is "very complicated" and backed by hundreds of research papers.