The Model Is Eating the Harness: Only Three Things Remain

The person who builds Cursor's harness put up a slide showing a minimal harness: nothing but shell access for the model, with all the API keys sitting next to it. He told the audience not to take photos, then asked: can something this bare-bones match a polished harness on coding benchmarks? Most people in the room thought not.

The Berkeley research says: yes.

The speaker is Jediah Katz, who joined Cursor in early 2025 and calls himself its most senior harness engineer. This year, following SpaceX's acquisition of Cursor, he moved to SpaceXAI. At the AI Engineer conference around October 20, he gave a two-sided answer to this question: for writing code, fancy harnesses can no longer make models smarter—but harnesses matter more than ever.

What Is a Harness, Exactly?

The term has only recently taken off. In 2023 the hottest job title was prompt engineer; by mid-2025 it became context engineer; earlier this year the industry settled on *agent harness*—referring to what tools like Cursor, Claude Code, and Codex actually do: a model wrapped in a shell.

Katz splits this shell into two layers. Beneath is the runtime: where the agent's main loop runs, where state lives, how work resumes after interruption. This layer resembles traditional software engineering—solid but not sexy. Above is the context and orchestration layer: which model to call when, what the model sees, how prompts and tools are configured, how memory gets compressed, plus a pile of tricks users never know about. When people casually say "harness," they usually mean this layer.

This shell used to be very busy. Earlier this year Cursor was showing off: a long-running multimodal harness orchestrating a large fleet of agents and machines to write a browser from scratch—so expensive that the demo crashed their own service once or twice.

Then the wind shifted.

The Evidence That It "Died"

The lineup declaring harnesses dead is impressive. Anthropic, at the same event, described deleting most of Claude Code's system prompt—the "about 80%" figure comes from Boris Cherny, creator of Claude Code, speaking at YC Startup School—with no measurable loss on coding benchmarks. YC's Garry Tan predicted harnesses will fade away, leaving only skills. Katz's own team has been deleting too: they once kept a small model dedicated to applying code edits to files, because frontier models of the time couldn't do it reliably; they also wrote extensive instructions teaching models how to explore codebases and self-check. Once models got smarter, all of it was deleted.

The hardest evidence is research. A study from Berkeley and Arena ran seven frontier models on a batch of harnesses including Pi, Codex, and Claude Code, finding that for newer, smarter models, swapping harnesses produced almost no difference on SWE-bench Light—but costs differed enormously, up to five times. Another study computed it more precisely: the same model cost twice as much in Claude Code as in Pi, with only a 1.1% score difference. Teams at Apple and EPFL reached the same conclusion on MLEBench: specialized MLE harnesses bring little improvement.

"Fancy harnesses no longer add coding intelligence today." Coming from Katz, that sentence carries different weight—he's the one who builds them.

The Data Flywheel

Why is this happening? Katz offered an explanation that is the conceptual core of the whole talk: the data flywheel.

Harnesses pull harder coding tasks into range, so more users can send agents to do that work; the trajectories of that work flow back to model labs as data, and models natively learn those skills. Once models learn them, harnesses delete the patches they once applied. One full turn of the flywheel: the harness gets thinner, the model gets stronger.

In other words, the harness is using its own work to eliminate its own work. Capabilities you take for granted today—exploring unfamiliar codebases, staying on track for hours at a time, running tests to self-verify—lived in harnesses just a few years ago, propped up by prompts and small models. Now they've moved into the weights.

Grok Bot Is Adding the Guardrails Back

The story doesn't end there. Intriguingly, while Cursor keeps deleting, SpaceXAI's Grok Bot keeps adding—new prompts, new guardrails.

The reason is simple. Grok Bot's harness isn't designed for coding; it manages an autonomous colleague: a long-lived coordination bot leading a team of working agents, like hiring a lawyer who personally manages strategy and delegates work to a team of paralegals. Making slides, booking travel, reporting and communicating, deciding when to ask a human versus act independently, remembering your work habits—models haven't cracked these tasks yet, and right now it's the harness that makes them feasible.

This completes the final link of the data flywheel: once a capability is commoditized into the model, it's the harness that opens the next frontier and pushes the flywheel again. Beyond coding, Katz gave the example of software factories: a bug comes into a channel, automation triages it immediately, a cloud agent picks it up to fix it, posts a before/after video as proof, and after deployment another agent watches production. One harness layer orchestrating a swarm of agents—he calls it a *meta harness*.

The Three Things Models Can't Eat

So as models keep getting smarter, what's left for the harness? Katz named three things—all outside the context-orchestration layer.

Where it runs. Agent work has stretched from minutes to days. A laptop that closes its lid can't sustain that; autonomous colleagues must run in the cloud. He offered a falsifiable prediction: within six months, ninety percent of agent work will not run on laptops.

What it can touch. A bare harness with all the keys can indeed solve problems, but nobody wants to wake up to a ten-thousand-dollar API bill, or a catastrophic irreversible operation. Permissions must narrow per person and per agent, and guardrails must be deterministic—this isn't a model capability problem, it's a trust problem.

How much it costs. This is where those two Berkeley studies truly teach their lesson: swapping harnesses barely moves the score, but costs can differ five-fold. The score belongs to the model; the bill belongs to the harness. Katz has spent the past month on exactly this: tuning cache hit rates, token efficiency, and model mix to cut Grok Bot's cost and latency without losing quality.

One detail deserves its own mention. The temporary guardrails that patch holes in the model are called shims internally—for example, if Grok Bot replies too slowly, a shim nudges it: "the user is still waiting." They log every shim, and record each time the model triggers one; when a shim stops being triggered, a bot automatically opens a PR to delete it. The flywheel isn't just a metaphor—they built a retirement mechanism into their own patches.

So is the harness dead? The part that made models write better code is indeed dead—consumed into model weights. But the part that opens new territory for agents, the part that manages trust, and the part that manages the bill are just getting started. In Katz's own words: yesterday's harness may be dead; long live tomorrow's.