The company Factory has released Droid 1.0, the most significant update to its coding agent harness, Factory Droid. The main change, now enabled by default, is programmatic tool calling: instead of chaining tool calls one by one, the model writes a program that executes its own tools, and a single request can trigger hundreds.
Factory calls a ‘harness’ the entirety of what surrounds the model: the prompts, the tools, the way their results come back, and the work loop. Droid 1.0 is model-independent. Users choose the one they want, or let Factory Router compose a blend deemed more cost-effective.
A program rather than successive calls
In a classic loop, each tool output is added to the model’s context, i.e., the text it must process, measured in tokens (the word fragments manipulated by language models). With programmatic tool calling, the model writes a program instead. Droid implements it as a tool named Script, written in JavaScript and executed in a sandbox (sandbox), an isolated environment that by itself has no access to files, the network, or the system.
Raw files, search results, and logs remain inside the program. Only the displayed output returns to the model, and independent calls execute in parallel. The approach builds on the CodeAct work (Wang et al., 2024).
Factory cites three examples from its own sessions. To fix a unit test, 15 tool calls required 5 model requests instead of 10, and the task completed in 41.6 s instead of 56.8 s. For cleaning up 151 tickets on Linear, 312 tool calls were needed; the Script used 2 requests and about 61,000 tokens, whereas direct calls, estimated by Factory, would require 36 and about 3.1 million. Finally, a scan of 5,384 tasks spread across 547 executions was done in 2 requests, and less than 0.1% of tool results reached the model’s context. On Claude Opus 5.5, Script mode reduces Droid’s cost by 15% compared to direct calls. It is enabled by default for ten models, including GPT-6.1 Sol and Kimi K3.
Results published by Factory
Factory compared each harness to Droid on the same model, with 3 trials per task, from October 23 to 28,2026. On the Terminal-Bench 4 benchmark, according to the company, Droid scores 65.2% with Claude Opus 5.5, ahead of Cursor and OpenCode (62.1%), Devin (59.6%), and Claude Code (57.6%). The announced cost is $4.73 per successful task. With GPT-6.1 Sol, Droid reaches 57.6% for $0.79 per successful task, compared to 54.0% for Codex, 51.0% for OpenCode, and 45.0% for Devin. Factory notes, however, that Droid takes 10% more time per successful task than Codex with this model, and that Cursor was not tested on GPT-6.1 Sol.
Other changes mainly concern lightness and responsiveness. The base prompt has gone from 20,887 to 13,055 tokens, a 37% reduction, and Droid now sends only 2 tool definitions at the start, down from 11 previously. According to Factory, a typical production request reads 98% of its inputs from the cache. Three identical calls trigger a retry, and six stop the loop. The application’s first display time drops from 1,839 ms to 488 ms median, and each additional session adds about 20 MB, compared to about 350 MB previously.
Droid 1.0 is offered as a desktop application for macOS and Windows, a terminal interface, and a Node SDK, @factory/droid-sdk. Paid plans start at $20 per month.
