# Droid 1.0: A New Era for Coding Agent Harnesses
## What Is Droid 1.0?
Droid 1.0 is a major release of an agent harness platform designed to work with multiple large language models. Unlike traditional frameworks that are tightly coupled to a single model, this platform lets developers choose any model they want — including ones accessed through their own API keys.
The defining feature of this release is that **programmatic tool calling is now enabled by default**. Instead of having the model trigger tools one by one in a standard loop, the model now writes a short JavaScript program that executes multiple tools in sequence or in parallel. The program runs inside a restricted sandbox that has no access to files, networks, or operating system resources. Only the printed output from that program flows back to the model for further reasoning.
This approach is inspired by recent research into agentic coding patterns and allows a single model request to drive hundreds of tool executions without repeatedly bloating the conversation context.
—
## Why Does This Matter?
In conventional agent loops, every tool result gets appended to the model’s input window. For tasks that require dozens or even thousands of tool calls, this quickly eats into context limits, increases costs, and slows everything down.
By letting the model write a program that handles tools internally, the harness dramatically reduces the amount of information that needs to be sent back and forth between the model and the tool environment. The model only sees the final printed results, not every intermediate step.
### Measured Impact
– **Unit test fixing**: A task that previously required 10 model calls and 15 tool interactions completed in just 5 model calls using the scripted approach — finishing in 41.6 seconds instead of 56.8 seconds.
– **Bulk issue management**: Moving and verifying 151 issues in a project management workflow required only 2 model requests and roughly 61,000 tokens via scripting, compared to 36 model requests and approximately 3.1 million tokens with traditional direct tool calls.
– **Continuous integration sweeps**: 5,384 individual jobs executed across 547 runs were completed in just 2 model requests, with fewer than 0.1% of tool results ever reaching the model’s context window.
—
## How Does the Program Sandbox Work?
The sandbox environment is intentionally locked down. It can execute JavaScript code and interact with the tools that have been loaded into the session, but it cannot read or write files outside its scope, make network requests, or access system-level resources.
This design choice provides several benefits:
1. **Security** — The model cannot accidentally or maliciously trigger operations beyond its intended scope.
2. **Predictability** — Tool execution happens in a controlled environment, making debugging and reproduction easier.
3. **Efficiency** — Intermediate data stays local to the script instead of being passed through the model context repeatedly.
Tool calls within the script can also run concurrently, which means independent operations — like searching multiple files or running several tests simultaneously — complete much faster than they would in a sequential loop.
—
## Performance and Benchmarking
The platform was evaluated on Terminal-Bench 4, a widely recognized benchmark for coding agent capabilities. Results vary by model, but Droid 1.0 consistently delivered top performance.
### Claude Opus 5.5 Results
| Harness | Pass Rate | Cost Per Passed Task |
|———|———–|———————|
| Droid 1.0 | 65.2% | $4.73 |
| OpenCode | 62.1% | Higher |
| Devin | 59.6% | Higher |
| Claude Code | 57.6% | Higher |
Droid achieved up to 24% lower cost and 16% to 26% faster completion times compared to other harnesses on the same model.
### GPT-6.1 Sol Results
| Harness | Pass Rate | Cost Per Passed Task |
|———|———–|———————|
| Droid 1.0 | 57.6% | $0.79 |
| Codex | 54.0% | Higher |
| OpenCode | 51.0% | — |
| Devin | 45.0% | — |
In this configuration, Droid was 10% slower per passed task than Codex but still delivered a higher pass rate at a lower cost per success.
—
## Other Key Changes in Version 1.0
Beyond scripted tool calling, the release includes a range of improvements driven by an internal automated system that analyzes failures and proposes optimizations across all supported model families.
– **Smaller initial requests**: The base prompt shrank by 37%, from roughly 20,887 tokens down to 13,055. Tool definitions shipped upfront dropped from 11 to just 2, with rarely used tools loading on demand.
– **Better caching**: About 98% of input in typical production requests now comes from cached content, since newly added tools are appended at the end rather than inserted into the shared prefix.
– **Smarter autonomy limits**: After three identical tool calls, the system nudges the model to reconsider. After six, it stops the loop entirely. Compaction summaries can now hold up to 16,000 tokens, double the previous limit.
– **Faster startup**: Time to first visual paint dropped from 1,839 milliseconds to 488 milliseconds. Extra sessions now share a single background process, adding only about 20 megabytes of memory each instead of roughly 350 megabytes.
—
## Getting Started
Droid 1.0 is available as a desktop application for both macOS and Windows, as a terminal interface, and through a Node.js software development kit called `@factory/sdk`. The SDK supports session management, streaming responses, and integration with Model Context Protocol (MCP) tools.
The software development kit itself is open source under the Apache 2.0 license, though the full platform’s licensing terms have not been publicly disclosed. Paid plans begin at $20 per month.
Developers can plug in any supported model or use Factory’s built-in routing system, which automatically selects cost-efficient model combinations based on the task at hand.
—
## Frequently Asked Questions
### What is an agent harness?
An agent harness is the surrounding infrastructure that manages how a language model interacts with external tools. It handles the prompt construction, tool execution, result feeding, and loop control — everything except the model itself.
### What is programmatic tool calling?
It is a pattern where the model writes a short program (in this case, JavaScript) that internally manages multiple tool calls. Instead of the harness running each tool one at a time and feeding results back to the model, the model bundles its tool needs into a single script that executes in a sandbox.
### Is the sandbox secure?
Yes. The sandbox has no access to the host file system, network, or operating system commands. It can only interact with tools explicitly provided by the harness and can only communicate back to the model through printed output.
### Does this work with my own API key?
Yes. The platform is model-independent, meaning you can bring your own API key for any supported model, including GPT, Claude, Kimi, and others.
### How does Droid compare to other harnesses on speed?
It depends on the model. On Claude Opus 5.5, Droid is 16% to 26% faster per passed task than competitors. On GPT-6.1 Sol, it is about 10% slower per passed task than Codex, but achieves a higher pass rate.
### What models is Droid 1.0 on by default for?
Script mode is enabled by default for 10 models, including GPT-6.1 Sol, Kimi K3, and Claude Opus 5.5.
### Is there a free tier?
The SDK is open source, and the platform offers paid plans starting at $20 per month. Pricing details for all features have not been fully disclosed.
—
## Conclusion
Droid 1.0 represents a significant step forward in how coding agent harnesses handle tool execution. By making programmatic tool calling the default — and shifting the responsibility of orchestrating multiple tool calls into a lightweight sandboxed script — the platform reduces context bloat, lowers costs, and improves pass rates on standardized benchmarks. The combination of leaner prompts, aggressive caching, and bounded autonomy controls makes it a compelling option for developers looking to integrate AI-powered coding agents into their workflows. Whether you use Claude, GPT, or any other supported model, Droid 1.0 offers a flexible, cost-efficient foundation for building and running autonomous coding tasks.
Thank you for reading
