Editorial illustration for Factory Droid 1.0 Uses Script Mode to Cut Unit Test Model Calls to 5
Factory Droid 1.0 Cuts Unit Test Calls by 95%
Factory Droid 1.0 Uses Script Mode to Cut Unit Test Model Calls to 5
Factory shipped Droid 1.0 this week, the biggest overhaul yet to its coding agent harness, and the headline change is something called programmatic tool calling. Instead of the model firing off one tool call, waiting for the result to land in context, then firing another, it now writes a JavaScript program that runs inside a sandbox and executes hundreds of tool calls in a single pass. Factory says this cuts a batch of unit tests down to just 5 model calls.
The numbers back up the pitch. Base prompt size dropped 37%, from 20,887 input tokens to 13,055. Tool definitions shipped up front fell from 11 to 2. Droid 1.0 runs on macOS, Windows, a terminal UI, and a new Node SDK called @factory/droid-sdk, and it works with any model a user wants to plug in, including their own API keys.
On Terminal-Bench 4, Factory's self-reported numbers put Droid at a 65.2% pass rate using Claude Opus 5.5, at $4.73 per passed task, beating every other harness tested on the same model. Not every pairing wins, though. With GPT-6.1 Sol, Droid runs 10% slower per passed task than Codex.
Factory has released Droid 1.0, the largest update yet to the Factory Droid agent harness. The main change is programmatic tool calling, now on by default. The model writes a program that runs its tools.
Why this matters
The headline number, 5 model calls instead of 10 for a 15-tool-call unit test fix, is a proxy for something we should watch closely: agent harnesses are becoming the place where real cost savings happen, not just the underlying models. Factory's Linear example makes the point sharper. 151 issues moved and verified in 312 tool calls using just 2 model requests and roughly 61,000 tokens is a batching win that no amount of prompt tweaking on a single model call would get you.
For developers and founders running agents at volume, this is the line item that actually shows up on the bill. We're skeptical of vendor-reported benchmarks from a single company's own sessions, and we'd want to see these numbers reproduced across providers before treating them as gospel. But the architectural idea, cut the base prompt 37% by shipping 2 tool definitions instead of 11 and let the model write a program rather than call tools one at a time, is a sound lever regardless of whose numbers you trust.
If script-mode style calling spreads to other harnesses, the real competition shifts from "which model is smartest" to "whose harness wastes the fewest tokens getting there."
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Box CEO Aaron Levie Says AI Will Drive "Biggest Cybersecurity Upgrade
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives Vercel Labs open-sources agent-desktop for model-agnostic computer use
From Archives Microsoft's TokenEconomizer Plugin Cuts Coding Costs by 30.9%
Common Questions Answered
What is programmatic tool calling in Factory Droid 1.0 and how does it differ from previous model calling methods?
Programmatic tool calling is a new feature in Factory Droid 1.0 where the model writes a JavaScript program that runs inside a sandbox and executes hundreds of tool calls in a single pass, rather than firing off one tool call, waiting for results, then firing another sequentially. This approach significantly reduces the number of model calls needed, cutting a batch of unit tests down to just 5 model calls instead of the previous 10.
How many model calls does Factory Droid 1.0 require to fix a 15-tool-call unit test compared to previous versions?
Factory Droid 1.0 reduces the model calls needed to fix a 15-tool-call unit test down to just 5 model calls, compared to 10 model calls in previous versions. This represents a 50% reduction in the number of times the model needs to be invoked for the same task.
What does the Linear example demonstrate about cost savings in agent harnesses?
The Linear example shows that 151 issues were moved and verified using just 2 model requests and roughly 61,000 tokens across 312 tool calls, demonstrating that significant cost savings come from batching tool calls at the agent harness level rather than from prompt tweaking on individual model calls. This illustrates that agent harnesses, not just the underlying models, are where real cost optimization happens.
Why does Factory Droid 1.0's Script Mode matter for AI coding agents?
Script Mode matters because it shifts the focus of cost optimization from the underlying models to the agent harness architecture itself, enabling substantial efficiency gains through batching. By allowing models to write programs that execute multiple tool calls in a single pass, Factory Droid 1.0 demonstrates that architectural improvements to how agents orchestrate tool usage can deliver more significant cost savings than model improvements alone.
Further Reading
- Droid 1.0: Building a Frontier Harness - Factory
- Factory's Droid 1.0 Beats OpenAI and Anthropic Agents While Cutting Costs 51x - AlphaSignal
- Programmatic Tool Calling - OpenAI
- Agent Harness - Microsoft Learn
- Best AI Coding Agents in 2026: Harness, Cost, and ... - Firecrawl
