Jediah
A senior engineer at Cursor and a core member of the Agent Harness team, focusing on research into agent frameworks and context orchestration.
Timeline (last 90 days)
Jediah believes some parts of the harness cannot be absorbed by models, including infrastructure and runtime (agent tasks range from minutes to days or longer and must run in the cloud), trust and security (requiring permission convergence and guardrails), and efficiency (both studies found significant per-task cost differences).
Jediah was an early engineer at Cursor and the person on the team who worked on the agent harness the longest, and recently joined SpaceX following Cursor's acquisition.
One of Jediah's reasons that the harness is not dead is that the harness unlocks new task frontiers for models, for example the Grokbot autonomous-colleague product, which covers various kinds of knowledge work beyond coding, orchestrating a group of worker agents with a long-lived coordinator bot.
Jediah lists three categories of harness components that models cannot absorb: infrastructure and runtime (agent tasks extending to hours or days require moving to the cloud), trust and safety (narrowing permissions and guardrails to prevent catastrophic operations or large API bills), and efficiency (studies all find significant differences in cost per task).
Jediah predicts that within six months 90% of agent work will run outside the laptop.
For the past month Jediah has focused on optimizing the efficiency of the GrokBot harness, and by adjusting cache hit rate, token efficiency and model mix, significantly reduced cost and improved speed without sacrificing quality.
Speaker Jediah is a senior engineer at Cursor, joined Cursor in early 2025, is the longest-tenured engineer on the agent harness team, and recently joined along with the company when it was acquired by SpaceX.
Jediah splits the harness into the runtime layer (controlling where the agent runs, loop execution, state storage to support task recovery) and the context and orchestration layer (deciding when to call which model and what context the model sees, including prompts, tools, memory, compaction and other mechanisms).
Jediah proposed the concept of a "data flywheel": the harness makes increasingly difficult coding tasks accessible, more users use agents to attempt these tasks, more coding data is shared with the labs, and once models natively learn these skills, the harness can delete what was originally added to make the tasks work.
All Facts (9)
Jediah believes some parts of the harness cannot be absorbed by models, including infrastructure and runtime (agent tasks range from minutes to days or longer and must run in the cloud), trust and security (requiring permission convergence and guardrails), and efficiency (both studies found significant per-task cost differences).
50%UnverifiedOne of Jediah's reasons that the harness is not dead is that the harness unlocks new task frontiers for models, for example the Grokbot autonomous-colleague product, which covers various kinds of knowledge work beyond coding, orchestrating a group of worker agents with a long-lived coordinator bot.
50%UnverifiedJediah was an early engineer at Cursor and the person on the team who worked on the agent harness the longest, and recently joined SpaceX following Cursor's acquisition.
50%UnverifiedJediah lists three categories of harness components that models cannot absorb: infrastructure and runtime (agent tasks extending to hours or days require moving to the cloud), trust and safety (narrowing permissions and guardrails to prevent catastrophic operations or large API bills), and efficiency (studies all find significant differences in cost per task).
50%UnverifiedFor the past month Jediah has focused on optimizing the efficiency of the GrokBot harness, and by adjusting cache hit rate, token efficiency and model mix, significantly reduced cost and improved speed without sacrificing quality.
50%UnverifiedSpeaker Jediah is a senior engineer at Cursor, joined Cursor in early 2025, is the longest-tenured engineer on the agent harness team, and recently joined along with the company when it was acquired by SpaceX.
50%UnverifiedJediah splits the harness into the runtime layer (controlling where the agent runs, loop execution, state storage to support task recovery) and the context and orchestration layer (deciding when to call which model and what context the model sees, including prompts, tools, memory, compaction and other mechanisms).
50%UnverifiedJediah proposed the concept of a "data flywheel": the harness makes increasingly difficult coding tasks accessible, more users use agents to attempt these tasks, more coding data is shared with the labs, and once models natively learn these skills, the harness can delete what was originally added to make the tasks work.
50%UnverifiedJediah predicts that within six months 90% of agent work will run outside the laptop.
50%