Factory ships Droid 1.0 with a Script mode to cut agent token use

Factory has released Droid 1.0, the largest update yet to the agent harness behind its enterprise coding product. The headline change is Script mode: instead of calling each tool directly in a long conversational loop, the agent writes a short program that invokes tools on its behalf inside a sandbox.

The problem it targets is familiar to anyone running long-horizon agents. In a standard loop, every tool call and its raw output is appended to the model's context, so instructions, tool definitions and history get re-sent turn after turn. That inflates token bills and slows work that spans hundreds of steps.

Script mode is an implementation of programmatic tool calling. The model writes JavaScript that runs with no file, network or system access of its own. A single script can make hundreds of tool calls within one model request, so instructions and history are sent once; printed output is capped in size and full results are logged for the model to search later.

Factory argues the approach buys reliability as well as efficiency. Loops, checks and branches spelled out in code run deterministically, reducing the chance that a model plans a step and then quietly drops it during execution. Independent calls can also run concurrently, so a batch takes roughly as long as its slowest call.

The company describes Droid as an agent harness: everything around the model, including prompts, tool definitions, how results come back and the loop that ties them together. Factory says Droid 1.0 outperforms the native harnesses built for individual models on the same tasks while using resources more efficiently.

Factory also stresses that the harness is model-independent by design, tuning prompts, tool descriptions and calling modes for each supported model. The argument is that tuning against a single model makes it hard to separate real progress from adaptations to that model's quirks, and that improvements which generalize across models are more durable for enterprises that switch providers.

The strategic bet is that coding-agent competition is shifting from model choice to the execution loop around the model. If agents increasingly run for hours or days, context hygiene, token efficiency and deterministic execution become the difference between a demo and a deployment. Making scripted tool calling the default is a wager that frontier models will increasingly act through code rather than one JSON call at a time.

What to watch is how Script mode holds up on messy enterprise repositories, and whether rival coding agents follow. The sandbox's permission boundaries, audit logs and failure recovery will matter as much as speed once security teams review the design.