Jediah Katz

Subway Stories brings his software work into a different setting. Betaworks describes it as an interactive visualization built with MTA Open Data, created with Julia Han and Mark Zitelli for the MTA’s 2024 Open Data Challenge. The project uses public transit data to explore how New Yorkers travel. His GitHub profile lists earlier work at Figma and an association with Anysphere. In his AI Engineer talk, he describes working on Cursor’s agent harness before moving to SpaceX.[1][2][3]

The software around the model

Katz describes the agent harness as the software that connects a model to tools, preserves state and supplies context. In his AI Engineer talk, he separates runtime duties, such as recovery after interruption, from prompts, memory and orchestration. He argues that stronger models can absorb behaviors previously supplied by the harness, including repository exploration and self-checking. The surrounding system then moves toward harder tasks.[2]

Durability, permissions and efficiency

Some responsibilities remain outside model weights. A model cannot itself supply persistent compute or enforce a user’s authority over external actions. Infrastructure, scoped access and cost control require surrounding software. Katz recommends treating behavioral interventions as removable machinery that product teams can retire as models improve.[2]