state, orchestrates work, applies permissions, verifies outputs and decides when a task is complete. This is not merely packaging: it determines the actions available to a model and the feedback loop through which it can correct itself.
The article reports that Claude Opus 4.6 built the same lightweight whiteboard app in both Factory Droid and Claude Code, with both passing five stated requirements. In that single run, Factory finished 11 minutes sooner and used roughly half as many tool calls; it also fixed a blocking hit-testing problem before reporting completion. The mechanism is: The harness selects context, executes tools, returns observations, maintains state, verifies work and determines termination; faster inference can make more iterations and checks practical, leaving tool-call orchestration and verification as constraints.
Why now: As inference speeds improve, multi-step work can become interactive, increasing the practical importance of the surrounding harness and making model-only comparisons less informative. The evidence boundary is: The article is vendor-authored and reports one task and one run; it does not establish that Factory Droid generally outperforms Claude Code or that the result generalizes across tasks. The first implication is: “Treat agent capability claims as properties of a model–harness configuration, and evaluate the whole loop.” The first watch condition is: “Look for repeatable, multi-task evaluations that report completion quality alongside tool calls, latency and cost.” Those tests would show whether the reported efficiency difference generalizes beyond this demonstration.