Factory releases Droid 1.0 with Programmatic Tool Calling and 37% smaller first-request prompt

Details

Factory released Droid 1.0, a model-independent agent harness featuring Programmatic Tool Calling (PTC) and a Script tool to improve efficiency and reliability.
On Terminal-Bench 4, Droid achieved a 65.2% pass rate with Claude Opus 5.5 at $4.73 per task, outperforming Cursor, OpenCode, Devin, and Claude Code by 3.0 to 7.6 points.
PTC allows agents to write JavaScript programs that execute in a sandbox to batch tool calls, reducing model requests and context window usage; one example cut input tokens from 3.1M to 61,000.
The first-request input token count was reduced by 37% from 20,887 to 13,055, with tool definitions shrinking by 45% through on-demand loading.
Prompt cache optimization appends new tools to the end of requests, achieving 98% input cache reuse in typical production scenarios.
Latency improvements include a median time to first paint drop from 1,839 ms to 488 ms and a 646 ms median time to first model request via parallelized startup tasks.
Droid 1.0 is available for macOS and Windows, with a TypeScript SDK supporting multi-turn sessions and custom MCP tools.

Key point