본문 바로가기
Opinion · 2026.10.19

Strip out the coding aids in your harness that the model has already learned, and invest in cloud runtime, permission scope, and cost per task — the score stays the same but the cost changes

Smarter Models Killed the Harness. Here's What's Left — Jediah Katz, SpaceXAI — AI Engineer · Read the original ↗

This page is a translation of the Korean original. Read the Korean original →

What you get

  • Harness code gets lighter as prompts and guardrails the model has absorbed are removed
  • Multi-day long-running tasks continue regardless of laptop sleep
  • Cost per task and response time go down at the same quality

When to use it

He says a harness is needed when extending agents beyond coding into knowledge work such as making slides, booking business trips, and email, and when building a "software factory" that chains multiple agents, such as automatic bug triage → fixing agent → post-deployment monitoring.

How to do it

  1. Remove the prompts and guardrails you added to patch coding intelligence, to the extent the model has absorbed them.
  2. Move the agent loop and state storage onto a cloud runtime rather than a laptop so long-running work can resume.
  3. Do not hand every API key to the model; narrow permissions to specific agents or people and put guardrails in place.
  4. Adjust cache hit rate, token efficiency, and model mix to lower cost per task and latency without hurting quality.
  5. Gather shims in one place and keep usage logs, and have it open a cleanup PR automatically when they stop firing.

Where I agree or hold back

The diagnosis that the harness cannot add intelligence in the coding domain is acceptable. It gives you grounds to check whether any instructions you use tell the model to do what it already does. However, predictions such as 90% churn within six months and the proposal to show users less of the internal workings lean on the speaker's own product context, so it is better to set them aside in your own environment.

Background — what the original covers

This is a talk by Jediah Katz, who built agent harnesses at Cursor longer than anyone, now speaking after joining SpaceX. As models got smarter, devices such as small models for code editing and codebase exploration instructions became unnecessary one after another, and studies have also reported that the choice of harness does not change coding scores much. He explains this as a cycle in which the harness makes itself obsolete, while saying that, where work outside coding opens up, as with Grokbot, he is putting guardrails back in.

#Strategy

← All practitioner tips