Don't Rush to Choose an Agent Framework—First Check Whether Your Harness Is Set Up
A comparison chart separates Agent framework and Agent harness: the former assembles workflows, while the latter configures an agent that can keep working. A three-day selection meeting is less useful than first checking whether tools, context, memory, and subagents are connected.
Akshay Pachaar posted a comparison chart that places Agent framework and Agent harness side by side. On the left side of the chart, in blue, is the framework: start, model adapter, branching, tools, custom steps, state, end. The flow is defined in advance, like assembling components into a line. On the right side, in green, is the harness: tools, files and shell, context, model, memory, subagents, plus your instructions, skills, tools, and data. The task is open-ended, and the agent keeps working. At the bottom of the chart are listed LangChain, CrewAI, Google ADK, OpenAI Agents SDK, as well as Hermes Agent, Codex, Deep Agents, and Claude Code. The original video comes from his tweet, reposted by Daily Dose of Data Science: link.
One point is worth putting up front: the framework solves how to draw the workflow; the harness solves what the agent relies on to keep working. Many teams spend three days in a selection meeting arguing LangChain vs. CrewAI, then go back to run a demo and find that tools are not connected, context is not managed, memory is not retained, and subagents are not configured. The workflow looks beautiful, but the agent cannot get work done. For people with a development background, the framework is close to the orchestration layer; for non-technical readers, you can think of it as a pipeline blueprint, while the harness is setting up a workstation for the agent, with tools, files, memory, and instructions at that workstation.
Aryxa wrote in a reply that LLMs are commodities; the orchestration loop around them is not. Viv mentioned a saying: if you are not the model, you are the harness. Sherif Zaidan added: You either die a model or live long enough to become a harness. Akshay replied: Exactly. These lines get the point across: models will change, and it is the tools, context, memory, subagents, and instructions in the harness that determine whether an agent can continuously complete tasks. Before choosing a framework, first check whether the harness is set up.
发布时间: 2026-11-02 21:32