Agent Harness: The Operating System for Agents

Anthropic’s engineering team argued in 2025: agent value is in the harness, not the model. Worth reading twice if you build AI products.

What Is a Harness?

Harness = the full runtime wrapping the LLM—managing perception, decision, execution, and feedback loops.

Analogy:

• LLM = CPU (inference)
• Harness = OS (scheduling, memory, I/O, process control)

Same GPT-4o, different harness → very different outcomes.

Core Components

┌─────────────────────────────────────────────┐
│                  Harness                     │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐  │
│  │ Context   │  │ Tool     │  │ Loop     │  │
│  │ Manager   │  │ Router   │  │ Controller│  │
│  └──────────┘  └──────────┘  └──────────┘  │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐  │
│  │ Skill     │  │ Sandbox  │  │ Eval &   │  │
│  │ Loader    │  │ / Safety │  │ Monitor  │  │
│  └──────────┘  └──────────┘  └──────────┘  │
│         ┌──────────┐                         │
│         │   LLM    │                         │
│         └──────────┘                         │
└─────────────────────────────────────────────┘

1. Context Manager

The bottleneck is often context, not raw model IQ.

Jobs:

• Choose what enters context (files, skills, MCP results, history)
• When to compress/summarize old turns
• Layout for LLM attention efficiency

Central battleground in 2026 AI engineering. See the context engineering article.

2. Tool Router

LLM emits “call tool_X” → harness:

• Parses tool calls
• Routes to MCP or built-ins
• Handles errors, retries, timeouts
• Formats results back into context

3. Loop Controller

Agents loop:

while not done and steps < max_steps:
    response = llm.call(context)
    if response.has_tool_calls:
        results = execute_tools(response.tool_calls)
        context.append(results)
    else:
        done = True
return response

Harness sets max steps, completion criteria, checkpoint/resume.

4. Skill Loader

Pick 1–3 relevant skills for the task—Cursor scans .cursor/skills/ and matches intent.

5. Sandbox / Safety

Agents edit files and run commands. Harness provides:

• Filesystem boundaries (workspace only)
• Command allowlists
• Network controls
• Human confirmation gates (delete, PR, etc.)

6. Eval & Monitor

Production harnesses track:

• Tool success/failure per step
• Token usage
• Task quality signals
• Anomalies (loops, privilege abuse)

Three Product Harnesses Compared

Component	Cursor	Claude Code	Devin
Context	index + open files + @ refs	whole repo + terminal	browser + IDE + terminal
Tools	MCP + built-ins	built-ins (+ MCP growing)	full stack
Loop	agent mode, interruptible	long autonomous runs	fully async
Sandbox	workspace	local shell	cloud VM
Skills	SKILL.md on demand	CLAUDE.md	internal behaviors

Philosophy:

• Cursor: human-in-the-loop collaboration
• Claude Code: long autonomous execution
• Devin: submit task, return later

Design Tradeoffs

Autonomy ←──────────────────────→ Control
Devin                          Cursor Tab
(fully autonomous)               (step confirm)

Context size ←──────────────────→ Cost/speed
Devin (whole repo)               Cursor (@file precision)

Generality ←──────────────────────→ Domain depth
General agent                    Code-specific harness

No universal best—match the scenario.

Minimal Custom Harness

class MinimalHarness:
    def run(self, goal: str, max_steps: int = 20):
        context = self.build_initial_context(goal)
        for step in range(max_steps):
            response = self.llm.chat(context, tools=self.available_tools)
            if response.tool_calls:
                for call in response.tool_calls:
                    result = self.execute_tool(call)
                    context.append(result)
            else:
                return response.text
        return "Max steps reached"

Simple surface—each method hides major engineering choices. That is the harness moat.

Summary

Everyone uses similar models. Harness differentiation explains why Cursor beats raw API calls. Next: the highest-ROI harness component—context engineering.