Eleven Coding Harnesses Share One Habit: They Do Not Import Agent Frameworks
A 2026 source-code study of eleven coding agents finds hand-rolled loops, not LangGraph, and skills ahead of MCP.
Direct answer
A 2026 source audit of eleven production coding harnesses found that none import a general agent framework and none retrieve code with vector embeddings.
Harness engineering only got its name in early 2026, but the runtimes behind coding agents are already old enough to compare. A source-code study of eleven production systems treats an agent as a model plus a harness: the loop, tools, context handling, safety checks, and extension points that connect the model to a repository.
The set is Claude Code, Codex CLI, Gemini CLI, Mistral Vibe, OpenHands, Aider, Mini-SWE-Agent, Hermes, Pi, OpenCode, and OpenClaw. The paper maps seven subsystems and records 13 observations that show up across vendors. It does not rank the products. It describes how they are built.
Frameworks are absent from the runtime
The sharpest finding is an absence. Across the Python, TypeScript, and Rust in that corpus, no harness imports a general-purpose agent framework such as LangChain, LangGraph, or AutoGen. Gemini CLI does not import Google's own agent frameworks either. The field is running on hand-written async loops.
Code lookup is just as plain. None of the eleven systems retrieve repository code with vector embeddings. They use deterministic tools: ripgrep, tree-sitter, glob, and Markdown context files the agent can discover on its own. For a coding harness, search is a filesystem problem before it is a retrieval-model problem.
Skills edged ahead of MCP
Extension standards have started to settle. SKILL.md-style skills appear in 9 of the 11 systems, and MCP in 8 of 11. The study also notes registries, trust tiers, and the first skills written by agents themselves. ACP is present in six systems and has picked up a hosting role: OpenHands can run Claude Code, Codex, or Gemini CLI as interchangeable backends.
For OctoLink's harness coverage, the practical reading is narrow. A harness is the body that lets one model finish work. Importing an orchestration framework is a separate choice, and the shipping coding agents in this audit have mostly declined it.
Source: Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents.
FAQ
- Which products were in the study?
- Claude Code, Codex CLI, Gemini CLI, Mistral Vibe, OpenHands, Aider, Mini-SWE-Agent, Hermes, Pi, OpenCode, and OpenClaw.
- Do these harnesses use LangGraph?
- The audit says none of them import LangChain, LangGraph, AutoGen, or similar general-purpose agent frameworks.