Factory Droid 1.0: The Model-Independent Coding Agent Harness That Rewrites the Rules of Tool Calling
Introduction:
Factory has released Droid 1.0, the largest update yet to the Factory Droid agent harness. The headline change is programmatic tool calling, now enabled by default: rather than issuing tools one at a time, the model writes a program that orchestrates its tools. As a result, a single model request can drive hundreds of tool calls.
TL;DR
- Size: The base prompt is now 13,055 input tokens, down 37% from 20,887. Only 2 tool definitions ship up front, down from 11.
- Runs on: A macOS and Windows desktop app, a terminal UI, and the Node SDK
@factory/droid-sdk. Works with any supported model, including your own API keys. - Performance: Highest pass rate of every harness Factory tested, on the same model, at lower cost.
- Best case: 65.2% on Terminal-Bench 4 with Claude Opus 5.5, at $4.73 per passed task (self-reported).
- Worst case: With GPT-6.1 Sol, Droid takes 10% longer per passed task than Codex.
What Is Droid 1.0?
Droid 1.0 is a model-independent coding agent harness from Factory. Factory defines a harness as everything around the model—the prompts, the tools, how tool results are returned, and the execution loop. Users can pick a model directly or let Factory Router handle model selection, and the harness adapts its behavior accordingly.
Programmatic Tool Calling and Script Mode
The central architectural shift in Droid 1.0 is programmatic tool calling. Instead of the conventional pattern—where the model requests one tool call, waits for the result, then requests the next—Droid 1.0 operates in what Factory calls Script Mode. The model generates a program that invokes tools, and that program executes the calls in sequence.
This approach delivers several advantages:
- Higher throughput per request: One model request can trigger hundreds of tool calls, reducing round-trip overhead.
- Lower token consumption: With only 2 tool definitions loaded up front (down from 11), the base prompt shrinks to 13,055 input tokens—a 37% reduction from the previous 20,887.
- Lower cost per task: Fewer model round-trips translate into measurable cost savings on complex engineering tasks.
Benchmarks and Real-World Performance
Factory reports that Droid 1.0 achieves the highest pass rate of every harness it tested, on the same model and at lower cost. On Terminal-Bench 4, Droid 1.0 with Claude Opus 5.5 reaches a 65.2% pass rate at $4.73 per passed task (self-reported).
The results are not uniformly dominant. With GPT-6.1 Sol, Droid takes 10% longer per passed task than Codex—a reminder that harness performance remains model-dependent even when the harness itself is model-independent.
Model Independence and Deployment Options
Droid 1.0 is designed to work with any supported model, and users can bring their own API keys. It ships across multiple surfaces:
- A macOS and Windows desktop app
- A terminal UI
- The Node SDK
@factory/droid-sdk
In 2026, as coding agents proliferate across IDEs, terminals, and CI pipelines, model independence has become a differentiator: teams want to swap models as capabilities and pricing shift, without rewriting their agent workflows.
Why It Matters
The harness layer—prompts, tool definitions, result handling, and the agent loop—is increasingly where competitive advantage in coding agents is decided. Droid 1.0's Script Mode suggests a broader industry trend: moving from turn-by-turn tool invocation toward programmatic orchestration, where the model writes code that drives its own tooling. If that pattern holds, expect more harnesses in 2026 to adopt similar approaches to reduce token overhead and increase per-request capability.
For teams evaluating coding agents, Droid 1.0 offers a compelling combination: a leaner prompt footprint, lower cost per passed task, and the flexibility to run on whichever frontier model delivers the best results at any given time.
via MarkTechPost