Former OpenAI researcher Andrej Karpathy has warned that the biggest current mistake in AI is pushing agents to do real work before developers have mastered "harness engineering" for the underlying model.
AI agents · Andrej Karpathy
Master the harness before forcing agents to work
Karpathy warns that the biggest current mistake in AI is pushing agents into real work before developers understand "harness engineering," the scaffolding that governs how a model uses tools, memory, prompts and feedback.
5 years
The time cost Karpathy attributes to the same mistake made at OpenAI in 2016
2016
The year Karpathy says the error was first made at OpenAI
The cost of skipping the groundwork
Each block is one year of cost Karpathy cites, in order.
Pattern: agents pushed into production before harness practices matured.
Karpathy ties the warning to his own experience but does not name the projects or teams involved.
What the harness covers
The surrounding system the article describes as the key variable alongside the model.
Model
→
Harness
Tools
Memory
Prompts
Feedback
→
Agent at real work
Open questions
The warning is brief, and key details are still missing.
How readiness for agent deployment would be measured
Which concrete harness practices are considered essential
Whether harness engineering becomes a formal discipline
Whether companies publish evaluation methods for autonomous readiness
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article...