What is an Agent Harness?

<aside> 🦥

Sources

The Anatomy of an Agent Harness – LangChain

Harness engineering for coding agent users – Martin Fowler

</aside>

🦥 Sloth's Simple Definition

<aside> 🦥

Agent = Model + Harness. The model is the brain (Claude, GPT, Gemini). The harness is everything else wrapped around it: the code, tools, memory, loops, and rules that turn a smart-but-helpless brain into something that actually gets work done.

The line to remember: "If you're not the model, you're the harness."

</aside>

Why This Is THE Concept of 2026

<aside> 🦥

Models are basically a commodity now: Claude, GPT, and Gemini all perform similarly. So what makes one agent way better than another? The harness.

One teardown of Claude Code found that roughly 98% of it is the harness, not the model. The intelligence is rented; the harness is the moat.

</aside>

What's Actually In a Harness

<aside> 🦥

System prompts & instruction files (CLAUDE.md, AGENTS.md, skill files, subagent prompts)
Tools and their descriptions (plus MCP servers)
The orchestration loop: the while-loop that keeps the model going until the task is done
Context management: compaction, memory, deciding what to load and when
Bundled infrastructure: filesystem, sandbox, browser
Guardrails & permissions: approvals before risky actions, error handling
Observability: logs, traces, cost/latency tracking </aside>

Inner Harness vs Outer Harness

<aside> 🦥

Inner harness = built by the tool maker (Claude Code's loop, prompts, tool dispatch). You don't touch this.
Outer harness = built by you: your rules, skills, hooks, subagents, and how you wire multiple agent sessions together. This is where your leverage is. </aside>

🦥 The Takeaway

<aside> 🦥

Stop obsessing over which model is smartest. The real skill in 2026, often called harness engineering, is building a great system around the model. Treat every mistake the agent makes as a chance to improve your harness, and it gets better every time the underlying model does. 🦥

</aside>