Factory has released Droid 1.0, a major upgrade to its AI coding-agent platform designed to make autonomous software development faster, more efficient and less dependent on any single artificial intelligence model. Announced on October 29, 2026, the release introduces a redesigned agent harness, with programmatic tool calling emerging as its most significant feature. The system allows an AI model to write JavaScript programs that execute multiple tool operations in a single workflow, reducing the need for repeated model interactions.
The update reflects a broader shift in AI-assisted software development. As coding agents take on increasingly complicated assignments involving large codebases, testing, debugging and project management, the infrastructure surrounding an AI model becomes almost as important as the model itself. Factory says Droid 1.0 is designed to address this challenge by improving how models access tools, manage context, reuse prompts and execute longer tasks. The company also says the platform works across supported models, giving developers more flexibility when choosing their AI systems.
What Is Factory Droid?
Droid is an AI coding agent developed by Factory. Rather than functioning only as a chatbot that generates code in response to prompts, it is designed to work through software-development tasks using tools and an execution environment.
The platform’s agent harness manages the instructions given to a model, the tools available to it, the way tool results are returned and the process through which the agent completes a task. These components influence how effectively an AI model can work on complicated projects.
Factory argues that coding agents increasingly operate across hundreds of interactions and may spend hours or days completing assignments. In conventional systems, instructions, tool definitions and conversation history can be repeatedly sent to the model. This increases processing requirements and can make long-running tasks more expensive.
Droid 1.0 attempts to reduce that overhead by changing how the agent interacts with its tools.
Programmatic Tool Calling Is the Biggest Change
The central feature in Droid 1.0 is programmatic tool calling, which Factory has enabled by default for supported models.
In a conventional coding-agent workflow, a model requests a tool operation, receives the result and then decides what to do next. It may need several additional model requests to complete a sequence of related operations.
With programmatic tool calling, the model can instead write a JavaScript program describing the entire workflow. Droid executes that program in a sandbox, allowing it to perform multiple tool calls and process their results before returning a summary to the model.
For example, an agent working on a large software project might need to:
- Read hundreds of files to identify a particular function.
- Search for related code and potential errors.
- Run tests and inspect the results.
- Update selected files.
- Verify that the changes have not introduced new problems.
Rather than receiving every intermediate result as a separate model interaction, the agent can execute much of this sequence through a single program.
How the Script system works
Factory implements this capability through a feature called Script mode. The model writes JavaScript, which runs in a controlled sandbox without independent file, network or system access. The script uses the tools provided by Droid to perform the requested operations.
This approach has three important advantages.
First, it reduces unnecessary information entering the model’s context. Raw file contents, logs and search results can remain inside the program, while the model receives only the information needed to make its next decision.
Second, it can reduce the number of model requests. A single script can perform numerous operations without asking the model to intervene after every tool call.
Third, it improves consistency in repetitive workflows. Loops, checks and verification steps can be executed through deterministic code rather than relying entirely on the model to remember every operation.
Factory says a script can perform hundreds of tool calls within one model request.
Factory Reports Improvements in Speed and Cost
Factory has published benchmark comparisons showing that Droid 1.0 performs competitively against other coding-agent environments, including Codex, Claude Code, Cursor, OpenCode and Devin.
The company tested the systems using Terminal-Bench 4, a benchmark designed to evaluate agents on terminal-based software engineering tasks. Its comparisons used the same underlying models across different agent harnesses.
According to Factory, Droid achieved the following results:
| Benchmark comparison | Factory’s reported result |
|---|---|
| GPT-6.1 Sol versus Codex | Droid scored 3.5 points higher and cost 12% less |
| Claude Opus 5.5 versus competing harnesses | Droid scored 3.0–7.6 points higher |
| Claude Opus 5.5 cost comparison | Up to 24% lower cost |
| Claude Opus 5.5 task completion time | 16%–26% less time per passed task than tested alternatives |
| GPT-6.1 Sol task completion time | 10% longer than Codex, despite 12% lower cost |
These figures are based on Factory’s own testing and should not be treated as independent confirmation that Droid will outperform competing products in every development environment.
The company also demonstrated a workflow in which 15 tool operations took 41.6 seconds using Script mode, compared with 56.8 seconds through conventional direct tool calls. The test involved the same underlying operations, but Script mode reduced the number of model requests from 10 to five.
Prompt Caching and Smaller Context Requirements
Another major part of Droid 1.0 focuses on reducing the amount of information that needs to be processed repeatedly.
AI models can reuse previously processed prompt information through prompt caching. However, changes near the beginning of a request can invalidate the cached portion that follows.
Factory says Droid now appends newly discovered tools to the end of the request rather than inserting them into an earlier tool list. This helps preserve the existing cached conversation context.
The company reports that 98% of input tokens in typical production requests are read from cache. It also says 98.6% of requests are strictly append-only, meaning the system did not identify changes that would break the cached prefix.
Droid’s initial prompt has also become smaller.
| Component | Earlier version | Newer version | Change |
|---|---|---|---|
| Tool definitions | 15,798 tokens | 8,528 tokens | -46% |
| Startup reminders | 3,463 tokens | 2,362 tokens | -32% |
| System prompt | 1,618 tokens | 2,155 tokens | +33% |
| Total initial prompt | 20,887 tokens | 13,055 tokens | -37% |
Factory says the reduction was achieved mainly by sending fewer tool definitions at startup. Instead of loading 11 definitions immediately, Droid now sends two and loads less frequently used tools when required.
A smaller initial prompt can reduce repeated processing, particularly in sessions involving many model interactions.
Faster Startup and Better Long-Running Sessions
Droid 1.0 also includes changes beyond the model interaction system.
Factory says it has improved application startup, session management and the handling of background tasks. The company reports that median time to first paint fell from 1,839 milliseconds to 488 milliseconds, while the time until the application was ready to accept a prompt declined from 1,876 milliseconds to 618 milliseconds.
The platform has also been redesigned to make multiple concurrent sessions more resource-efficient. Factory says that previously, separate sessions could require separate processes using approximately 350 MB of memory each. Under the new architecture, sessions share a daemon process, with each additional session adding approximately 20 MB.
For longer-running assignments, Droid can move commands and subagent work into the background. It can then resume the agent when the operation finishes instead of repeatedly checking whether the task has completed.
The system also includes safeguards for stalled or repetitive activity. Factory describes mechanisms that detect repeated tool calls, incomplete responses and tasks that stop making progress.
Availability and Developer Access
Factory says Droid 1.0 is available through its desktop application for macOS and Windows, as well as a terminal interface.
Developers can also access the same agent harness through the Droid SDK, distributed as @factory/droid-sdk. The SDK allows applications to start Droid sessions, stream agent activity and provide custom tools through the Model Context Protocol (MCP).
The model-independent approach is another important part of the release. Developers can select supported models or use Factory Router, which is designed to choose an appropriate model for a task.
This gives development teams more flexibility to change models as new systems become available without necessarily rebuilding their entire coding-agent workflow.
The Bigger Picture
The release highlights how competition in AI coding is expanding beyond the underlying language models. As models become more capable of writing and modifying software, the systems that coordinate tools, manage context and verify results are becoming increasingly important.
Factory is betting that a model-independent harness can offer developers a consistent environment while allowing them to use different AI models. Its programmatic tool-calling approach also points towards a future in which agents execute larger workflows through code rather than relying on repeated conversational instructions.
However, benchmark improvements do not automatically translate into better results for every developer. Real-world performance will depend on the programming language, repository size, task complexity, model selection, integrations and the quality of the verification process.
Looking Ahead
Factory Droid 1.0 represents a significant step in the development of autonomous coding agents. Its focus on programmatic tool calling, prompt caching and more efficient session management addresses some of the practical limitations that emerge when AI agents work on large software projects. The ability to use different supported models through one harness could also make it attractive to teams that do not want to depend on a single AI provider.
The next challenge will be demonstrating that these improvements translate into reliable performance across everyday development environments. As competitors continue improving their own coding agents, the ability to complete complex assignments accurately, securely and at a predictable cost is likely to become a defining factor in the market.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
