The Three Camps of Open-Source AI Harnesses

foxgem

Harness is one of the hottest topics in the agent world right now. Most models are good enough these days; what really sets agents apart is the layer wrapped around the model.

Setting vendors’ own harnesses aside, the open-source world has quietly split into three Camps.

The LangChain Camp: langchain → langgraph → deepagents

The veteran. It started as a component library, grew into orchestration, and has now arrived at the batteries-included Deep Agents.

  • Language: Python first, with JS/TS versions keeping pace.
  • Ecosystem: the whole LangChain suite, plus LangSmith for observability and evals.

The Pi Camp: pi → oh-my-pi → flue

Everything revolves around pi, a minimal core, with the community building on top. oh-my-pi is a fork that goes the batteries-included route as a terminal coding agent. Flue, from the Astro team, uses pi as its foundation and turns it into a programmable, headless agent framework.

  • Language: TypeScript.
  • Ecosystem: community-driven, with a small core and plenty of extensions; Flue already runs on Cloudflare.

The AI SDK Camp: ai sdk → workflow → eve

Made by Vercel. The AI SDK handles model and tool calls, the Workflow SDK handles durable execution, and eve builds a Next.js-style agent framework on top.

  • Language: TypeScript.
  • Ecosystem: the Vercel platform, with AI Gateway, Sandbox and other infrastructure ready out of the box.

At a Glance

LangChain CampPi CampAI SDK Camp
Main languagePython (plus JS/TS)TypeScriptTypeScript
FoundationLangChainpiAI SDK
Orchestration / durabilityLangGraphLeft to the layers aboveWorkflow SDK
Out-of-the-box harnessDeep Agentsoh-my-pi, Flueeve
Backed byLangChain (the company)Community + Astro teamVercel
DeploymentAnywhere, paired with LangSmithAnywhere; Flue runs on CloudflareVercel first

Finally

The three start from very different places: an established Python framework, a minimal community core, and a cloud platform. Yet their paths look strikingly alike: foundation → orchestration → an out-of-the-box harness.

The harness is shifting from a feature to the product itself. As for which one to pick and how to put it to use, I’ll leave that for you to ponder.