After Reading Four Million Lines of Code, Not One of 11 Programming Agents Uses LangChain

An 83-page paper did a labor-intensive job: four authors opened the source code of 11 production-grade programming agents, read about four million lines of Python, TypeScript, and Rust, and tried to answer one question: what exactly is an Agent.

The paper's answer is very short: Agent = Model + Harness. The model is rented; the Harness is the runtime layer outside the model. It is responsible for connecting the LLM to the real world: sending prompts, reading tool calls, managing context, controlling permissions, and deciding which things run automatically.

These 11 systems are Claude Code, Codex CLI, Gemini CLI, Mistral Vibe, OpenHands, Aider, Mini-SWE-Agent, Hermes, Pi, OpenCode, and OpenClaw. The paper breaks the Harness into 7 parts: loop, model bridge, tools, memory and context, security, orchestration, and extension. The minimal Harness can even do without sub-agent orchestration.

The most striking thing is two absences. Across four million lines of code, not a single runtime imports general-purpose agent frameworks such as LangChain, LangGraph, or AutoGen. Nor does any system use vector embeddings to retrieve code; everyone uses ripgrep, glob, and markdown files. Reading code down to this conclusion is a bit absurd: on one side are framework marketing pages, on the other are hand-written asynchronous loops and deterministic search in production systems.

There is also an adoption data point: SKILL.md appears 9 times across the 11 systems, and MCP 8 times. The way to extend is changing too; configuration and skill files are more common than grand plugin ecosystems.

In the post, slash1s compresses the paper into a few points and casually inserts a demo of Minara Harness, saying it points the same 7 parts at the market. Looking through the paper's list, Minara is not among the 11 systems. Some commenters also pointed this out. When reading, keep the research sample and the product demo separate.

The paper abstract confirms that the authors gave 18 design recommendations and a 90-line minimum viable Harness scaffold. The tweet summary says 90 lines cover 10 of those recommendations. For people who want to get hands-on, this threshold is not high; for people who mystify Agents, this number is not very friendly. Writing a loop is not hard. What's hard is deciding how to trim context, how to tighten permissions, and how to provide tools.

The paper's argument is that in the first half of 2026, programming Harnesses changed from tools into platforms. Ordinary users can also take away a judgment from this: with the same model, different Agent products differ greatly in experience, and the difference is often in the Harness. For readers with a development background, the 7 parts and two absences are more useful than framework marketing pages.

Paper: https://arxiv.org/abs/2609.00006

Original post: https://x.com/slash1sol/status/2119397444349784100

(The image is from the original post)