Factory launches Droid 1.0: programmatic tool calling by default, hundreds of tool calls per model request, leads native harnesses on its own Terminal-Bench 4 runs

Background and pain point

Coding agents now run for hours across hundreds of turns, and every turn re-sends instructions, tool definitions and history. In a classic loop the model calls one or two tools at a time and every output (whole files, search results, logs) goes back into context, so long tasks get more expensive, slower and more likely to drop steps. Factory released Droid 1.0 on Oct 29, calling it the largest update yet to its agent harness, meaning everything around the model: prompts, tools, how results come back, and the loop.

Architecture and mechanism

The headline change makes programmatic tool calling (PTC) the default through a Script tool. The model writes JavaScript that runs in a sandbox with no file, network or system access of its own, making up to hundreds of tool calls in one model request, running independent calls concurrently, returning only the program output to context and logging full results for later search. Factory''s example: moving and verifying 151 Linear issues took 312 tool calls, about 36 requests and 3.1M input tokens with direct calls, versus 2 requests and 61K tokens as a Script. Other changes: newly loaded MCP tools are appended to the end of the request to keep the prompt cache intact (a typical production request reads 98% of input from cache); the first request shrank from 20,887 to 13,055 tokens (-37%), with up-front tool definitions cut from 11 to 2; the loop detects stalls, repeats and dropped tasks; and compaction summaries grew from 8,000 to 16,000 tokens. Factory also describes an internal Agent Factory where agents study benchmark failures and real-session friction, propose changes and test them on every model family before engineers decide what ships.

Benchmarks

All numbers are Factory''s own runs (all Terminal-Bench 4 tasks, Oct 23-28, three trials per task, high reasoning). GPT-6.1 Sol: Droid 57.6% ($0.79 per passed task), Codex 54.0%, OpenCode 51.0%, Devin 45.0%; Factory says Droid beats Codex by 3.5 points at 12% lower cost but takes 10% longer. Claude Opus 5.5: Droid 65.2% ($4.73), Cursor and OpenCode 62.1%, Devin 59.6%, Claude Code 57.6%, with 16% to 26% less time per passed task. The methodology notes adjustments: some Opus 5.5 trials stopped by content moderation were rerun on Opus 5, Devin tasks where it stopped to ask were counted as passed if Codex passed, and Cursor tokens lost on two reconnects were estimated from transcripts. These are vendor results with no third-party reproduction yet.

Getting started

Droid 1.0 is available now as a desktop app for macOS and Windows or as a terminal UI. To build on the same harness, use the Droid SDK (npm package @factory/droid-sdk), which runs the same Droid as the CLI: start Droid from Node, hold multi-turn sessions, stream its work, add custom tools over MCP, or connect to a running Droid daemon. Droid is model-independent, with bring-your-own keys or the Factory Router choosing models. It suits tool-heavy long tasks such as reading many files, triaging tickets in bulk or staged database migrations; compare cost and pass rate on your own codebase rather than relying on vendor charts.