Factory makes coding agents call hundreds of tools from one script
Droid 1.0 turns programmatic tool calling on by default and claims better Terminal-Bench results at lower cost.
Factory has released Droid 1.0, a rebuilt harness for its coding agent that makes programmatic tool calling the default. Instead of asking the model to invoke tools one by one, Droid can have it write a JavaScript program that calls hundreds of tools inside a sandbox.
The company says this design keeps raw files, search results and logs out of the model’s context, reduces repeated model requests and lets deterministic loops verify every step. Independent tool calls can run concurrently. In one example, moving and checking 151 Linear issues required 312 tool calls; Factory says Script mode completed the work in two model requests and about 61,000 input tokens, versus an estimated 36 requests and 3.1 million tokens with direct calls.
Factory also published Terminal-Bench 4 comparisons using the same models across different agent harnesses. With GPT-6.1 Sol, Droid scored 57.6% at $0.79 per passed task, versus 54.0% for Codex. With Claude Opus 5.5, Droid scored 65.2%, ahead of Cursor and OpenCode at 62.1% and Claude Code at 57.6%.
Those numbers come from Factory’s own runs and have not been independently verified. Harness configuration can materially affect both quality and cost, which is also the company’s main argument: the model is only one part of an effective coding agent.
Droid remains model-independent and supports users’ own API keys. An automatic router can select models, while Factory says its internal evaluation system tests harness changes across multiple model families before release.
