Droid 1.0: Building a Frontier Harness | Factory
Factory announces Droid 1.0, an agent harness that lets models write JavaScript programs to run many tool calls at once instead of calling tools one at a time. They say it cuts cost and context use while improving results across several frontier models.
- Script mode lets the model make hundreds of tool calls in one request; on Claude Opus 5.5 it cut Droid's cost by 15% versus direct calls.
- One example cleanup moved 151 Linear issues (312 tool calls) in 2 model requests and about 61,000 tokens, where direct calls were estimated at 36 requests and roughly 3.1M input tokens.
- Droid is tuned per model and claims up to 20% better efficiency than native harnesses on the same tasks, with the gains coming from keeping raw tool output inside the program rather than the model's context.
Factory's Droid 1.0 is an agent harness, meaning everything around the model: prompts, tool definitions, how tool results come back, and the loop that drives it all. The headline claim is that on the same tasks with the same models, Droid scores higher at lower cost than every other harness they tested, and outperforms the native harnesses built for those models by up to 20% on efficiency. The biggest change behind that is programmatic tool calling (PTC), where the agent writes a program that invokes tools rather than calling them one at a time. Factory ships this as a Script tool: the model writes JavaScript that runs in a sandbox with no direct file, network, or system access, and can make hundreds of tool calls in a single model request. Independent calls run concurrently, and only the printed output re-enters the model's context.
The reasoning for making this the default is fairly concrete. Raw file contents, search results, and logs stay as variables inside the program, so the model's context stays clean. Instructions and history get sent once per Script instead of once per call, which cuts request count. And loops and checks run deterministically in code, so the model can't quietly skip a case it planned to handle. Factory's examples make the scale visible. A Linear cleanup moving 151 issues across 312 tool calls took 2 model requests and about 61,000 input tokens, where they estimate direct calling would have taken 36 requests and roughly 3.1 million input tokens. A CI sweep checking 5,384 desktop jobs across 547 runs used two requests, and the model saw only a five-number summary, with under 0.1% of tool results reaching its context. Against direct calls on Claude Opus 5.5, Script mode cuts cost by 15% overall, with most of the savings concentrated in the 12.5% of Scripts that make six or more calls.
The article is also candid about where the design choices matter and what they're built on. Factory lists a few decisions that mattered most: giving programs real structured data rather than strings, passing file contents and commands as raw input blocks to avoid escaping overhead and malformed calls, and returning short summaries from edits so files aren't pasted back into context. The harness is model-independent by design, with prompts and tool descriptions tuned per model and an auto router called Factory Router that picks among models. Their tuning process, which they call the Agent Factory, has agents study benchmark failures alongside real user friction and test each change across every affected model family before engineers decide what ships. Script mode is on by default for a long list of GPT and Claude models plus Kimi K3, so the vendor-neutral pitch is the core of the product. Worth reading if you're building or buying agent infrastructure, though the numbers are Factory's own, measured on their tasks, and the cost comparisons rely partly on estimates for the direct-call baselines.