Wink Pings

OpenAI Co-founder: When Agents Fail, Don't Rush to Blame the Model

OpenAI co-founder Andrej Karpathy says in a video that if your Agent keeps failing, the problem is very likely not the model but the Harness. He breaks down a core chain—model, harness, loops, agent—and uses lessons from OpenAI back then as an example.

Andrej Karpathy recently posted a 5-minute video on X about AI Agents. He opened with a blunt point: everyone is forcing Agents to do work, but no one is willing to first thoroughly understand the Harness. OpenAI made this mistake back then too, and it cost them several years.

Karpathy outlined a core chain: Model → Harness → Loops → an Agent that can actually get work done. He put it directly: when an Agent fails, it has nothing to do with the model itself. Most teams spend every week swapping models, always hoping the next new model will save them with one click; the top 1% players polish the Harness once and for all, and the whole system immediately runs smoothly.

What is a Harness? You can think of it as the scaffolding outside the model: how prompts are organized, how tools are called, how context is managed, how errors are handled, how evaluations are run. The model is the engine; the Harness is the chassis. No matter how powerful the engine is, if the chassis doesn't match, it will still flip on the road.

In these 5 minutes, Karpathy mentions using Loops and Graphs to build the Harness. Loops handle the cycle, letting the Agent rotate through observation, action, and feedback; Graphs handle process orchestration, clearly laying out branches and jumps. These concepts need to be implemented in practice—not just sounded clever.

In the comments, a netizen said that when writing scripts with Cursor, you first need to feed the rules file and context clearly, and only then does the work stabilize; swapping models actually doesn't help that much. Someone else asked whether Harness specifically refers to environment interaction or the evaluation framework—actually, it includes both. Another commenter put it even more bluntly: if the framework is poorly built, no matter how brilliant the model is, it will run like an idiot.

Of course, there are different opinions: saying that Agent failures have absolutely nothing to do with the model is a bit absolute. That's true—model capability is still the foundation. But Karpathy's reminder is worth hearing: many teams spend time swapping models instead of solving the Harness that is truly holding them back. Sharpening the axe does not delay the chopping—this still holds in Agent development.

发布时间: 2026-11-23 04:06