312 actions, two requests: Droid 1.0 rethinks how AI agents work

Factory evolves Droid with programmable tool execution, reducing requests, costs, and delays on complex tasks.

Moving 151 tickets in Linear, verifying each change, and executing 312 tool calls: an operation that previously required up to 36 requests to the model can now be completed in just two. This demonstration summarizes the evolution of Droid 1.0, the new agentic runtime environment developed by Factory. Rather than improving the artificial intelligence model itself, the company is intervening in how it organizes and executes its actions. This approach is intended for complex software projects, where successive interactions with tools can quickly multiply costs and delays. Hundreds of Actions in a Single Request

The major evolution relies on Programmatic Tool Calling (PTC), now used by default with several compatible models. In traditional operation, an agent calls its tools sequentially, retrieves their responses, and then queries the model again to determine the next step. Each interaction enriches the conversation history and increases the information to be processed.

Droid's new Script mode alters this setup. The model writes a JavaScript program capable of orchestrating several hundred tool calls within a single execution. Operations can be chained, conditioned on previous results, or run simultaneously when independent. Intermediate data remains within the program, while the model retrieves only the information necessary to continue its work. The code executes in an isolated environment, without direct access to files, the network, or the system outside of authorized tools.

In the example of the 151 Linear tickets, all 312 calls remain necessary to move and inspect the items. What decreases are the exchanges with the model: Factory estimates that a traditional execution would have consumed about 3.1 million input tokens, compared to 61,000 with Script. These figures come from an internal benchmark in which the traditional scenario is modeled, rather than a direct measurement of two identical executions.

The same method was used to inspect 5,384 continuous integration tasks spread across 547 runs. Through 550 tool calls, the agent identified 41 failures while passing only a five-value summary back to the model. Less than 0.1% of the raw results thus entered its working context. Performance Compared to Codex and Claude Code

Evaluations conducted on Terminal-Bench 4 allow Factory to measure the differences between Droid and several competing environments, notably Codex, Claude Code, Cursor, Devin, and OpenCode. Tests were carried out from October 23 to 28, 2026, using the same models and three trials per task.

With GPT-6.1 Sol, Droid achieves a 57.6% success rate, compared to 54% for Codex, at a 12% lower cost on compared tasks. With Claude Opus 5.5, it reaches 65.2% with an average cost of $4.73 per successful task. According to results published by Factory, cost savings can reach up to 24% against certain competing environments.

Speed, however, varies depending on the configuration. On Claude Opus 5.5, Droid completes successful tasks 26% faster than Claude Code and uses 17% fewer tokens. With GPT-6.1 Sol, it consumes 12% fewer tokens than Codex, but requires 10% more time. These measurements cover tasks completed successfully by both compared environments under Factory's protocol, and do not constitute an independent assessment of their overall performance. An Agent Factory to Improve Agents

Behind these evolutions lies an internal infrastructure named Agent Factory. It automates part of Droid's optimization work by analyzing failures encountered during benchmarks and difficulties observed in real-world sessions. Agents identify recurring issues, propose modifications, and submit them to controlled experiments across different model families.

Engineers retain final say over the changes integrated into the product. This method notably makes it possible to adapt instructions, tool descriptions, and execution mechanisms without relying on a single provider. Droid can thus operate with several proprietary or open-weight models, while Factory Router automatically selects a configuration based on needs.

Other optimizations concern context management. Between versions 0.170.0 and 0.250.0, the initial instructions and information passed to the model were reduced from 20,887 to 13,055 tokens, representing a 37% decrease. Rarely used tools are now loaded on demand, and cache retention limits the reprocessing of previously transmitted information. Factory states that 98% of input tokens in a typical request are read from cache, based on its production telemetry. An Agent Independent of the Model Powering It

The improvement also extends to the user experience. Between the August and October releases, the median time required to display the interface dropped from 1,839 to 488 milliseconds. Session startup, handling multiple concurrent agents, and closing the application were also optimized.

Droid 1.0 is available on macOS and Windows, as well as via a terminal interface. Developers can also integrate its capabilities into their applications using the Droid SDK, which notably enables session management, execution tracking, and tool connectivity via the Model Context Protocol.

This evolution illustrates an important distinction in AI agent development: the capabilities of the model and the efficiency of the system operating it do not necessarily advance at the same pace. Factory's results show that modifying the execution environment can improve performance, cut costs, and enhance the handling of complex tasks without training a new model. The true extent of these gains, however, remains to be confirmed beyond the benchmarks and use cases published by the company.