Is the Agent Harness Dead? Jediah Katz, Head of Grok Bot & Cursor Harness, Argues: Coding patches are being absorbed by models, but opening up new tasks, safety, and efficiency still depend on the Harness!
Models are getting smarter, and the part of the Harness that "helps the model write good code" is indeed disappearing. But the Harness still has two responsibilities that won't disappear: first, pushing the model to open up new task boundaries; second, providing infrastructure, trust and safety, and efficiency that the model cannot absorb.
YouTube video:
https://youtu.be/7Y4sbZgBdnc?is=9PAd3si0khM5XkOJ
Speaker and Background
@jediahkatz joined Cursor in early 2025, and is the engineer who has worked on the Agent Harness the longest. After Cursor was acquired, he joined SpaceX and took part in the Grok Bot Harness. He traced the field's changing terminology: in 2023 "prompt engineer" was popular, in mid-2025 it became "context engineer," and this year the industry finally settled on "Harness engineering."
What the Harness Is Made Of
The Harness is the model's "eyes and hands." Upward, it connects to the App users use; downward, it connects to the computer and third-party services (such as MCP). He divides it into two layers:
Runtime layer: decides where the Agent loop runs and how state is saved so it can resume after an interruption. This is closer to traditional software engineering.
Context and orchestration layer: decides when to call which model and what the model can see, including prompts, tools, memory, compression, and various behavior-tuning techniques. He describes this layer as the "Wild West."
Why Some Say the Harness Is Dead
Early this year was the peak of the Harness: Cursor used a long-running Harness called Grind Swarm to build a browser from scratch. The cost was extremely high, and it reportedly brought Cursor down once or twice.
A few months later, the wind shifted to "the Harness is dead": Anthropic's Tariq said Claude Code deleted most of its system prompt; YC's Garry Tan argued that from now on only Skills matter; Cursor itself removed many guardrails, such as the small model dedicated to applying code edits, and coding instructions like "how to explore the codebase and how to check your own work."
Research evidence: UC Berkeley tested 7 frontier models with different Harnesses (Pi, Codex, Claude Code), and scores on SWE-bench Lite showed no significant difference; Apple and EPFL found that dedicated MLE Harnesses brought no clear improvement on MLE-Bench either.
His judgment: today's fancy Harnesses no longer increase the model's coding ability. The reason is a data flywheel: the Harness makes harder tasks feasible, so more users use Agents for those tasks, more data flows to model vendors, models internalize those skills, and the Harness can delete the patches it originally added. In a sense, the Harness is making itself obsolete.
Why the Harness Is Not Dead
Opening new boundaries: after coding ability is "commoditized" by models, the Harness opens up the next class of tasks, and the flywheel turns another round. He takes his own Grok Bot as an example: at Cursor, they kept deleting prompts and guardrails, but at Grok Bot they added many back, because web workflows like making slides and booking travel, reporting progress without disturbing the user, judging whether to act directly or ask first (for example, before sending an email on the user's behalf), and learning the user's habits through memory are things models have not yet fully mastered. Another example is the "software factory": bug reports come in and are triaged automatically, a cloud Agent fixes them and records before-and-after comparison videos, and after release another Agent watches the production environment. This amounts to a "meta-Harness" orchestrating multiple Agents.
What models cannot absorb:
Infrastructure: Agent tasks have grown from minutes to hours and days, so they cannot run on a laptop that goes to sleep; they must run in the cloud.
Trust and safety: permissions are narrowed per Agent and per person, preventing catastrophic operations, or waking up to find a $10,000 API bill.
Efficiency: although the two studies above found small differences in scores, the cost per task differs significantly. Over the past month he significantly reduced cost and latency without quality dropping, by improving cache hit rates, token efficiency, and optimizing model combinations.
Trends Ahead
Products are running ahead of models again, and the Harness will become messy before it becomes clean again, so the design must make it easy to ship and retire features. One manifestation is exposing fewer internal mechanisms to users, the reasoning being Hyrum's Law (anything you expose will be depended on by users), at the cost of reduced transparency.
Designing Agent-oriented primitives: for example, "subscriptions," using a single unified tool that lets an Agent respond to Slack messages, CI failures, schedules, and various other events, instead of building a separate integration for each event type.
Centrally tracking "shims": for known bad habits of models, first apply temporary patches with deterministic rules (such as reminding it "the user is waiting for you"), recording each time one fires; once a patch no longer fires, a Bot automatically opens a PR to delete it.
Moving up the abstraction level: the previous generation of Harnesses was composed of multiple LLM calls, while the new generation takes the Agent as the basic unit.
His prediction: within 6 months, 90% of Agent work will move off personal computers and run in the cloud.