Factory Droid 1.0 makes coding agents write programs to run tools
Coding agents just stopped asking for tools one by one. Factory Droid 1.0 lets them write a single script to execute hundreds of tasks.
Tools · Source: Factory
What happened
Factory just released Droid 1.0. It is a model-independent coding agent harness. The biggest update is programmatic tool calling. It is now turned on by default for ten models. Instead of a model calling tools one by one, it writes a JavaScript program to do the work. This changes how agents handle long-horizon tasks.
This code runs inside a secure sandbox. The sandbox has no direct file, network, or system access of its own. The model uses a Script tool to execute the code. One model request can now drive hundreds of tool calls. In one test, Droid moved and verified 151 Linear issues using 312 tool calls in just two model requests. Only the printed output goes back to the model context.
Factory claims this beats every other harness on the market. On Terminal-Bench 4 using Claude Opus 5.5, Droid scored 65.2 percent. It beat Cursor, Devin, and Claude Code. With GPT-6.1 Sol, it scored 57.6 percent. Factory says the base prompt also shrank 37 percent down to 13,055 tokens. The app latency for first paint dropped from 1839 milliseconds to 488 milliseconds.
Key facts
- 13,055 — Tokens in the new base prompt, down 37 percent.
- 65.2% — Self-reported pass rate on Terminal-Bench 4 with Claude Opus 5.5.
- $4.73 — Cost per passed task using Claude Opus 5.5.
- 98% — Input read from cache in a typical production request.
- $20 — Starting monthly price for paid plans.
Why it matters
The old way of building agents clogs the context window. Every tool output gets dumped back into the prompt. Programmatic tool calling changes the architecture entirely. The agent reasons upfront and writes a workflow. Raw files, search results, and logs stay inside the program as variables. You save massive amounts of tokens and cut down on expensive model requests. Independent calls even run concurrently. The batch takes only as long as its slowest call.
This shifts the competitive moat from the underlying model to the harness surrounding it. Factory tunes prompts and tool descriptions across multiple models. They use an internal Agent Factory to study benchmark failures alongside real user friction. Agents propose changes and test them on every affected model family. When you optimize the loop around the model, you get better results than relying on native model capabilities alone. Cheaper runs mean you can deploy more agents on the exact same budget.
For builders
Cut token costs with script mode
Direct tool calls eat your budget. Script mode cuts Droid cost by 15 percent on Claude Opus 5.5. Founders pay less for complex tasks like massive CI sweeps or tricky database migrations.
Bring your own API keys
Droid is model-independent by design. You can plug in your own API keys for models like GPT-6.1 Sol or Claude Opus 5.5. You control the model mix while Factory handles the routing and harness optimization.
Beware of self-reported benchmarks
Factory ran these Terminal-Bench 4 tests internally. Independent parties have not verified the 65.2 percent pass rate. Engineers should test the Node SDK against their own codebases before ripping out existing tools.
My take
I always say the harness is the real product, not the model. Coding agents are finally moving past the slow chat interface. Letting the AI write a deterministic program to do the work is how we actually scale these tools reliably.
Original reporting: Factory. This is my rewrite and opinion.