Factory Droid 1.0 Turns One Model Request Into Hundreds of Tool Calls

Factory has released Droid 1.0, the largest update yet to its coding agent harness. The main change is simple to describe but broad in effect: programmatic tool calling is now on by default.

Instead of sending every tool result back into the model’s context, Droid lets the model write a program that runs its tools. One model request can now drive hundreds of tool calls, with only the output the program prints returning to the model.

How Droid’s Script mode works

Factory describes Droid 1.0 as a model-independent coding agent harness. In Factory’s terms, the harness is everything around the model. Users can pick a model or let Factory Router choose a cost-efficient mix, while Droid handles the software that turns that model into a coding worker.

In a standard tool loop, every tool output lands in the model’s context. Droid’s programmatic tool calling, or PTC, changes that flow through a tool called Script. The model writes JavaScript, and that code runs in a sandbox with no file, network, or system access of its own.

Raw files, search results, and logs stay inside the program. Only printed output returns to the model, and independent calls run concurrently. The idea builds on CodeAct, described by Wang et al. in 2024.

Factory’s own sessions show the difference. In one unit test fix, 15 tool calls took 5 model calls instead of 10, and the task finished in 41.6 seconds instead of 56.8 seconds. In a linear cleanup, Script moved and verified 151 issues through 312 tool calls using 2 requests and about 61,000 tokens. Factory says its models used 36 direct requests and about 3.1M tokens for the same work.

Another CI sweep covered 5,384 jobs across 547 runs in 2 requests. Under 0.1% of tool results reached the model’s context. On Claude Opus 5.5, Script mode cuts Droid’s cost by 15% compared with direct tool calling.

Lower prompts, smaller requests, and faster sessions

Droid’s base prompt now uses 13,055 input tokens, down 37% from 20,887. Only 2 tool definitions ship up front, down from 11, which gives the first request a leaner starting point.

Newly loaded tools are appended to the end of a request so the typical production request reads 98% of its input from cache. Rarely used tools load on demand, and tool definitions fell 46% as part of the change.

Droid also adds limits for repeated actions. Three identical tool calls trigger a nudge, while 6 stop the loop. Compaction summaries now hold 16,000 tokens, up from 8,000.

The app changes reach beyond model requests. First paint fell from 1,839 milliseconds to 488 milliseconds, and each extra session now adds about 20 MB instead of 350 MB. Factory says most of these changes came from an internal Agent Factory loop, where agents read failures and propose changes, those changes are tested on every affected model family, and engineers review and merge them.

Factory’s Terminal-Bench 4 results

Factory tested each harness against Droid on the same model, with 3 trials per task from October 23 to 28, 2026. Only tasks that both harnesses passed in all 3 trials counted toward cost and time.

With Claude Opus 5.5, Droid scored 65.2% on Terminal-Bench 4, the highest pass rate of every harness Factory tested on that model. Factory reports a cost of $4.73 per passed task, and says Droid costs up to 24% less. Cursor and OpenCode scored 62.1%, Devin scored 59.6%, and Claude Code scored 57.6%.

On Opus 5.5, Droid took 16% to 26% less time per passed task and used 17% fewer tokens than Claude Code. With GPT-6.1 Sol, Droid scored 57.6% at $0.79 per passed task. Codex scored 54.0%, OpenCode 51.0%, and Devin 45.0%.

The comparison is not a clean sweep. Factory’s worst case came with GPT-6.1 Sol, where Droid took 10% longer per passed task than Codex. Factory’s results are self-reported.

Where developers can use Droid 1.0

Droid runs on the macOS and Windows desktop app, a terminal UI, and the Node SDK @factory/droid-sdk. The SDK runs the same harness from Node and includes sessions, streaming, and MCP tools. Droid works with any supported model, including a developer’s own API keys.

Programmatic tool calling is on by default for 10 models, including GPT-6.1 Sol and Kimi K3. Droid 1.0 ships as a desktop app and terminal UI, while paid plans start at $20 per month.

The larger idea behind the release matches a recent study that examined 11 coding agents. The study found that every system used the same seven parts, with the size of each part changing from one agent to another. Its simple conclusion was that an agent is a model plus a harness: the model does the thinking, and the harness is the software around it.

Droid 1.0 puts that harness in the spotlight. Its default Script tool changes how model requests handle large batches of work, while its prompt, caching, session, and loop controls target the cost and time around each task.

Stay connected via Google News

Follow us for the latest AI News