Agent Labs: Workload-Harness Fit

Agent labs such as Cursor, Intercom, Cognition, and Decagon have recently been launching their own vertical models, pursuing a vertical integration strategy to reduce dependence on large model providers and protect margins.

The technical resource allocation of agent labs splits into two directions: one camp focuses on directly training and updating model weights, while the other focuses on building agent engineering—context management, tool use, and other elements—or the optimal Harness.

The key to this choice lies in Workload-Harness Fit. Workloads vary in character according to four dimensions.

• Volume: The frequency with which a task is executed. It determines the generation speed of the data flywheel and the economic incentive for cost reduction.
• Value per execution: The economic impact of completing a task. It determines the scale of investment in model quality and the tolerance for error.
• Verification Properties: This refers to the Veracity, Proliferation, and Asymmetry of the reward signal.
• Time horizon: The complexity of sequential decision-making and tool interactions. The longer the horizon, the more exponentially difficult reward attribution and training-data generation become.