Why “Agent Development” Misleads: Framework vs. Harness in LLM Agents

Not Either‑or

Both LangChain users and Claude Code users claim to build an “Agent”, but they are actually operating at different architectural layers: one assembles components (Framework), the other runs a pre‑packaged solution (Harness).

What Framework Feels Like

Frameworks such as LangChain, CrewAI, and LlamaIndex give you building blocks—memory modules, tool interfaces, and orchestration logic. You must decide which memory backend to use (Redis or in‑memory), how to register tools, how to handle failures, and how to coordinate multiple agents. The framework defines the shape of the blocks but leaves the assembly to you.

The target audience is developers who want to construct agents; they need to understand concepts like Chain , Runnable , and tool schemas. Consequently, the learning curve is steep, and troubleshooting can involve swapping memory implementations, adjusting tool schemas, or reordering execution steps.

What Harness Feels Like

Harnesses such as OpenClaw, Claude Code, Cursor, Deep Agents, and similar products provide a complete, ready‑to‑run agent. After supplying an API key and a task description, the system handles memory, context management, tool calls, security checks, and state persistence automatically.

You only adjust high‑level parameters—changing the model, adding a tool, or tweaking the prompt. The underlying architecture, error‑recovery logic, and context handling are fixed by the harness creator.

Core Difference: Who Makes Decisions

Frameworks expose decisions to the user (memory choice, tool design, execution order), while Harnesses embed those decisions. This impacts what you can modify when problems arise: with a framework you can replace the memory layer, rewrite tool schemas, or change orchestration; with a harness you are limited to prompt engineering, model swapping, or adding/removing tools.

Boundaries Blur

The line between framework and harness is becoming fuzzy. Historically, web development moved from raw PHP/SQL to frameworks like Laravel, then to turnkey platforms like WordPress. Similarly, LLM agent tooling is following the same trajectory.

LangChain’s stack illustrates this: the core library is a Framework, LangGraph acts as a runtime, and Deep Agents sit on top as a Harness.

Which to Choose Depends on Your Problem

If an existing harness covers your task (e.g., Claude Code for programming, Deep Research for research, OpenClaw for automation), start with the harness to avoid reinventing memory management and error handling.

If your team lacks the expertise to maintain a complex framework, a harness reduces maintenance overhead. Conversely, if you need frequent model swaps, custom memory backends, or fine‑grained control, a framework’s flexibility is advantageous.

When requirements are unclear, begin with a bare API call to identify bottlenecks before deciding between framework and harness. In some cases, a simple ReAct loop (reasoning + acting) may be the cleanest solution.

Why Harness Appears Now

One year ago, limited options meant most developers used frameworks. Recent advances in model reliability—stable function calling, long context windows, and robust instruction following—have enabled the emergence of mature harnesses.

These harnesses now span programming, research, and task automation, but they still rely on the underlying model’s capabilities. As models improve, more harnesses become viable.

Understanding whether you are using a framework or a harness directly influences debugging, extensibility, and tool evaluation.