Models vs Harnesses: How Agent Frameworks Work

The most important architectural concept in agent orchestration is the separation between the model and the harness. Understanding this distinction is the foundation for everything you will build in this course.

The Brain and the Body

Think of it this way:

• The Model (Brain): The LLM that generates text, reasons through problems, and understands language. This is Claude, GPT, Gemini, a local model like Llama, or any other language model.
• The Harness (Body): The surrounding system that gives the model tools, memory, identity, communication channels, and the ability to take real-world actions.

The model is powerful but blind. Without a harness, it cannot read your emails, check your calendar, post to social media, or remember what you talked about yesterday. The harness provides all of these capabilities.

┌─────────────────────────────────────────┐
│              Agent Harness              │
│  ┌─────────┐  ┌─────────┐  ┌────────┐  │
│  │ Memory  │  │  Tools  │  │Identity│  │
│  └────┬────┘  └────┬────┘  └───┬────┘  │
│       │            │           │        │
│  ┌────▼────────────▼───────────▼────┐   │
│  │          LLM Model (Brain)       │   │
│  │    Claude / GPT / Gemini / Local │   │
│  └──────────────────────────────────┘   │
│  ┌─────────┐  ┌─────────┐  ┌────────┐  │
│  │Telegram │  │ Discord │  │ Email  │   │
│  └─────────┘  └─────────┘  └────────┘  │
│           Communication Channels        │
└─────────────────────────────────────────┘

Model-Agnostic Design

A well-designed harness is model-agnostic — you can swap the underlying LLM without changing the rest of the system. This is a core design principle in frameworks like OpenClaw.

Why does this matter?

Scenario	Best Model Choice
Complex reasoning and analysis	Claude or GPT-4 class models
Fast responses with low latency	Smaller local models
Cost-sensitive batch operations	Open-source models via API
Image understanding tasks	Multimodal models (Gemini, GPT-4 with vision)
Privacy-critical operations	Fully local models (Llama, Mistral)

With a model-agnostic harness, you can route different tasks to different models based on the requirements. The harness handles the complexity of switching between providers.

What the Harness Manages

The harness handles everything the model cannot do on its own:

1. Context Assembly Before each model call, the harness assembles the context: who the user is, what happened in previous conversations, what tools are available, and what the agent's identity and rules are.

2. Tool Execution When the model decides it needs to call a tool (send an email, search the web, read a file), the harness executes that tool and feeds the result back to the model.

3. Memory Persistence After each interaction, the harness saves important information to persistent storage so the model can access it in future sessions.

4. Channel Routing Messages come in from different channels (Telegram, Discord, email). The harness normalizes these into a common format, processes them through the model, and routes responses back through the correct channel.

The Partner System

One practical pattern in agent frameworks is the partner system — using a second, smaller model to debug and validate the primary model's output. If the primary model generates a response that seems off (too long, wrong format, potentially hallucinated), the partner model can flag it before the response reaches the user.

This is not redundancy — it is quality control. The partner model acts as a proofreader, catching issues before they become visible.

Key takeaway: The model provides raw intelligence. The harness provides everything else — memory, tools, channels, identity, and security. A great harness can make a modest model highly capable, while a poor harness wastes even the best model's potential.

Next: The shift from traditional workflows to agentic workflows — and what changes in how you work. :::