Meet Factory Droid 1.0: A Model-Independent Coding Agent Harness Built on Script Mode

Factory has released Droid 1.0, a major update to its coding agent harness, the prompts, tools and loop that sit around whichever model a team picks. The headline change is that programmatic tool calling is now on by default. Instead of feeding every tool result back to the model, Droid has the model write JavaScript that runs its tools inside a sandbox and returns only what the program prints.

Factory says that one model request can now drive hundreds of tool calls. In one of its examples, a cleanup of 151 project-tracker issues took two requests and roughly 61,000 tokens, against an estimated 36 requests and about 3.1 million tokens the old way. It also reports the top pass rate on Terminal-Bench 4 among harnesses it tested, 65.2% with Claude Opus 5.5.

If those savings hold, the cost of running agents on long, tedious jobs drops sharply, and the choice of harness becomes a real lever, not an afterthought.

We think Factory is right about the core idea. Pouring every log line and search result into a model's context is wasteful and distracting. Letting the model write a small program, keep the raw data inside it, and surface only the answer is how a careful engineer would work. The approach builds on published research, CodeAct, and Factory's own numbers point the same way.

Those numbers are also Factory's own. Every Terminal-Bench 4 figure comes from Factory's runs of competitors' harnesses, over a few days in late October. The methodology is more careful than most, with three trials per task and cost counted only on tasks both harnesses passed, but vendors running rivals' tools is not the same as an independent leaderboard. Factory deserves credit for printing a loss: with GPT-6.1 Sol, Droid is about 10% slower per passed task than Codex.

The risk we want named is oversight. When one request can fire hundreds of actions, a human watching the transcript sees a script and a summary, not each step. Factory's sandbox has no file, network or system access of its own, and the harness now nudges after three identical calls and stops after six. Those are good guardrails. But a script that moves 151 tickets correctly can move 151 tickets wrongly just as fast, and less of the evidence reaches anyone's eyes.

There is also a quieter detail: many of these harness changes were proposed by agents reading failures, then reviewed and merged by engineers. That is a sensible loop, and it is also how the tools that write software increasingly shape themselves.

Our read: a meaningful step toward cheaper, faster agents, with the bill for that speed paid in visibility. I'm watching for independent benchmark runs and for audit trails that show what each script actually touched.

Coding agent harnesses compete with each other and with model makers' own tools, such as Claude Code and Codex. Factory sells Droid as model-independent, supports customers' own API keys and offers a router that mixes models. It ships a desktop app, a terminal UI, and a Node SDK under Apache 2.0, with paid plans starting at $20 a month.