Factory Launches Droid 1.0: Its Coding Agent Writes Programs to Chain Hundreds of Tool Calls
Factory introduced Droid 1.0, the latest version of its coding agent's “harness.” The main change: the model writes a small JavaScript program that handles tool calling on its own, reducing the number of requests and context consumed. Factory claims higher scores than Codex, Claude Code, or Cursor on Terminal-Bench 4 at a lower cost, based on self-reported benchmarks.
On Thursday, October 29, at 11:31 PM Paris time, startup Factory introduced Droid 1.0, the most significant update to date for its coding agent’s engine. Factory does not build frontier models: its product is what the industry calls a “harness,” meaning everything surrounding the model (instructions, tools, result parsing, and the execution loop). The company positions it as more accurate, more cost-effective, and faster than agents designed directly by the model developers themselves.
The model writes a program instead of calling tools one by one
The centerpiece of Droid 1.0 is programmatic tool calling, now enabled by default. In a standard agent loop, the model calls a tool (such as reading a file or running a shell command), receives the output back into its context window, and then decides on the next step: each step requires a new API request, which repeatedly resends instructions and conversation history.
With the Script tool, the model instead writes a small JavaScript program that chains these calls itself. This script runs inside a sandbox with no native access to files, the network, or the operating system; it can interact only through Droid’s exposed tools. Raw intermediate data (files, logs, search outputs) remains stored in program variables, and the model only reads whatever the script outputs. Factory highlights three benefits: cleaner context windows, fewer API requests, and deterministically executed verification steps with no risk of the model “forgetting” an edge case along the way.
The company shared two examples from internal sessions. Cleaning up Linear issues (151 tickets updated and then verified, spanning 312 tool calls) required 2 requests and approximately 61,000 tokens, compared to an estimated 36 requests and 3.1 million tokens via direct tool calling. Reviewing 5,384 CI jobs was completed in 2 requests, with the model only viewing a five-number summary. Script mode is turned on by default for flagship models from OpenAI and Anthropic (including GPT-6.1 Sol and Claude Opus 5.5), as well as Kimi K3.
Factory also streamlined the initial prompt for each session, bringing it down from 20,887 to 13,055 tokens (−37%), primarily because Droid now supplies only 2 initial tool definitions instead of 11, loading remaining tools on demand. Lastly, the company introduced loop-prevention guardrails (such as terminating execution after six identical tool calls) and significantly reduced application startup times.
Results measured by Factory
On Terminal-Bench 4, Factory benchmarked Droid against rival harnesses using identical models. Paired with GPT-6.1 Sol, Droid reportedly scored 3.5 points higher than Codex at 12% lower cost, albeit taking 10% more time. With Claude Opus 5.5, it achieved scores 3.0 to 7.6 points higher than Cursor, OpenCode, Devin, and Claude Code, at up to 24% lower cost. Cursor was omitted from the GPT-6.1 Sol comparison, with Factory noting that Cursor lacks access to OpenAI’s models.
These figures stem entirely from Factory and have not been independently verified. The testing methodology, detailed at the bottom of its post, includes several notable caveats: cost and duration metrics only include tasks solved successfully by both harnesses across all three attempts, tasks triggered by Anthropic’s moderation filters were rerun on Claude Opus 5, and an instance where Devin paused to prompt the user was marked as a pass if Codex completed the task.
Getting started
Droid 1.0 is available immediately as a desktop application for macOS and Windows, as well as a terminal interface. To build custom tooling on top of the same execution engine, Factory provides a TypeScript SDK, @factory/droid-sdk, which enables running Droid inside Node environments, handling multi-turn sessions, and integrating custom tools via MCP.
Our take
Droid 1.0 exemplifies a broader trend across coding agents: performance gains are no longer solely driven by raw model capability, but by how the model is orchestrated. Having the model write a script to coordinate tool calling is not unprecedented (Factory itself references prior work by CodeAct, Cloudflare, Anthropic, and Codex), but establishing it as the default mode in a commercial, model-agnostic agent is a clear bet on frontier models’ ability to plan ahead. The cost argument is compelling: on long-running tasks, fewer network round trips and lighter context footprints directly lower the bill. That said, comparisons against Codex and Claude Code remain internal benchmarks with debatable methodological choices; independent evaluations will be required to confirm these claims.
Sources
- Factory (X) — Droid 1.0 announcement Primary source
- Factory — Droid 1.0: Building a Frontier Harness Primary source
- Factory — Droid TypeScript SDK documentation Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
