Smarter Models Killed the Harness. Here's What's Left — Jediah Katz, SpaceXAI

Highlights

• Jediah Katz joined Cursor in early 2025 as the longest-tenured engineer working on the agent harness before moving to SpaceX via acquisition.

• Studies from UC Berkeley and Apple with EPFL found that sophisticated harnesses no longer significantly change scores on coding benchmarks like Sweet Bench Lite and MLE bench for newer frontier models.

• The data flywheel process allows models to natively absorb skills once harnesses bring harder tasks within reach and generate high volumes of coding data for labs.

• GrokBot uses a multi-agent harness comprising long-lived coordinator bots that orchestrate fleets of worker agents for general knowledge work.

• Cloud execution is projected to run 90% of agent work within six months as tasks extend from minutes to days beyond local laptops.

Timeline

1. 00:00:00 - 04:30:00

   Evolution and the Question of the Dead Harness

   The terminology surrounding AI engineering shifted rapidly from prompt engineering in 2023 to context engineering and eventually harness engineering by early 2026. The harness itself functions as the primary bridge between the model and external environments, split cleanly between a deterministic runtime and a dynamic orchestration layer. Peak harness development culminated in projects like GrindSwarm, a multi-model architecture used to build an entire browser from scratch that required massive computing resources.

   • Prompt engineering evolved into context engineering by mid-2025, which then matured into agent harness engineering.
   • An agent harness wraps around a model to act as its eyes and hands, connecting it to computers and third-party services like MCPs.
   • The harness consists of a runtime foundation for state storage and an orchestration layer for managing prompts, tools, and memory.
2. 04:30:00 - 07:15:00

   Why People Say the Harness is Dead

   Industry figures from Anthropic, Y Combinator, and Cursor noted the removal of extensive system prompts and guardrails because frontier models no longer require synthetic scaffolding for tasks like codebase exploration and test execution. Empirical research confirms that simple setups perform comparably on coding benchmarks to highly complex specialized harnesses. This obsolescence is driven by a continuous data flywheel where harnesses enable harder work, generate training data, and ultimately transfer those capabilities directly into model weights.

   • Recent evaluations from UC Berkeley and Apple with EPFL demonstrate that advanced harnesses do not significantly boost benchmark scores on top models.
   • Frontier models have natively absorbed coding capabilities that previously required heavy guardrails and micro-model patches in the harness.
   • The data flywheel makes specific harness features obsolete as models learn complex skills directly from data generated during agent-assisted tasks.
3. 07:15:00 - 14:18:00

   Why the Harness Still Matters

   Despite native model improvements, the harness remains indispensable for expanding into new domains like autonomous slide deck creation, calendar management, and communication workflows. Furthermore, critical elements such as security permissions, API cost controls, and robust cloud runtimes cannot be absorbed by the model alone. Comparative cost analyses show that while scores remain identical across different harnesses, the underlying system architecture dictates the financial viability and operational speed of running tasks at scale.

   • GrokBot employs long-lived coordinator bots to manage worker fleets, requiring re-introduced prompts and guardrails for non-coding knowledge work.
   • Harnesses provide the underlying cloud infrastructure required for agents that must run continuously for hours or days without depending on a local laptop.
   • Infrastructure optimization through cache hit rate adjustments, token management, and model mixing significantly reduces task costs and improves execution speed.
4. 14:18:00 - 18:06:00

   Future Trends for Agent Harnesses

   The rapid pace of product development requires flexible primitives, such as unified event subscription frameworks that allow agents to react to Slack messages or CI failures without custom integrations. Temporary shims are logged centrally to track when model capabilities catch up so deprecated code paths can be automatically removed via pull requests. Ultimately, the future relies on higher-level agentic abstractions and cloud-native environments capable of supporting fully autonomous multi-step execution.

   • Product iteration speed is outpacing model releases, leading engineers to build modular subscription primitives and centralized shims for cleaner codebases.
   • Harnesses are shifting abstractions from individual LLM call compositions to multi-agent architectures acting as primary building blocks.
   • Ninety percent of agent workloads are expected to migrate off local laptops and into cloud infrastructure within six months.