- Jediah Katz argued that models learn the tricks a harness once supplied, so a fancy harness adds little coding skill.
- He said what remains is infrastructure, trust and security, and efficiency, including scoped permissions and guardrails against catastrophic actions or a $10,000 API fee.
Jediah Katz, a longtime agent-harness engineer at Cursor and now at SpaceX, made a pointed argument on the AI Engineer show. A harness is the software around an AI model that lets it read files, run tools and act. Katz said models are learning the clever tricks harnesses used to supply. What lasts, in his view, is the part a model cannot be trained to do: limit what an agent may touch and spend.
What was said
Katz said the harness makes itself obsolete through what he calls the data flywheel. Skills that harnesses once patched in get taught to the models themselves. His conclusion was blunt: "A fancy harness in this day and age no longer adds coding intelligence."
He then listed what models cannot absorb: infrastructure, trust and security, and efficiency. On security, he noted that a simple setup with all the API keys in a file can solve coding tasks. But he warned that "you probably don't want to give all your API keys in a file to the model." He said companies want permissions scoped to specific agents and specific people. Guardrails also matter for anyone who wants to avoid a catastrophic action or a $10,000 API fee.
On infrastructure, he said agent tasks have grown from minutes to hours to days. That cannot run on a laptop that may go to sleep, so the work has to run in the cloud. On efficiency, he cited two studies. Harness choice made no significant difference to benchmark scores, but it did make a significant difference to cost per task.
Why it matters
Our reading: the model's skill is increasingly something you rent, while the limits around it are something you must set yourself. If an agent holds broad keys and has no spend cap, the person who handed over the keys and the budget is the one who pays when it goes wrong. Katz did not frame it as a question of blame. But his list suggests practical questions for buyers. Does each agent get narrow permissions of its own? Is there a ceiling on spend? Is cost per task measured, given that the studies he cited found cost varied even when scores did not?
The other side
Katz builds harnesses for a living, so he has a stake in the claim that harnesses still matter. He also gave no figures on how often agents fail or how well guardrails prevent it. The $10,000 fee was an illustration, not a measured loss.
His cloud argument was about reliability over long tasks. The excerpts do not show him arguing that cloud isolation limits damage, so that link is ours. He also expected harnesses to get messier before they get cleaned up, so the picture is still shifting.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: Smarter Models Killed the Harness. Here's What's Left — Jediah Katz, SpaceXAI (2026-10-19)
