Factory Releases Droid 1.0 Establishing Programmatic Tool Calling as Default Standard

SAN FRANCISCO — Challenging the conventional turn-by-turn design of autonomous coding software, AI software harness pioneer Factory has officially launched Droid 1.0, a major architectural release that elevates Programmatic Tool Calling (PTC) to the default execution standard. Packaged as a model-independent developer harness available as both a terminal UI and a native desktop environment, Droid 1.0 shifts coding agents away from chat-like conversational tool loops, enabling frontier models to author and execute complex multi-step automation scripts within a single inference roundtrip. The resulting architecture achieved a record-breaking 65.2% score on Terminal-Bench 4 using Claude Opus 5.5 and 57.6% with GPT-6.1 Sol while slashing context window bloat by up to 60%.

Why Turn-by-Turn Tool Calling Breaks on Complex Repositories

Traditional coding assistants like Cursor Composer or early LangChain agent loops operate via rigid conversational turns: the model emits a JSON tool call (e.g., readFile("src/auth.ts")), the host environment responds with the file text, the model reads the file, decides it needs a second file, emits readFile("src/db.ts"), and so on. In repositories containing hundreds of microservices, an agent routinely burns 30 to 50 roundtrips just to map import dependencies.

Every single step reinjects the entire conversation history back into the LLM's context window. By step 20, input prompt caching costs skyrocket, context limits are breached, and the model begins hallucinating earlier instructions. Droid 1.0 breaks this paradigm by introducing Programmatic Tool Calling (PTC). Instead of issuing individual requests, the model writes a unified Python or Bash script that executes inside Droid's local sandbox.

# 1. Initialize Droid inside your repository
droid init

# 2. Launch autonomous refactoring with script-mode PTC enabled
droid run --ptc --model claude-opus-5-5 \
  "Refactor all authentication middlewares to use JWT RS256 with unit tests"

# Droid authors an executable sandbox script that reads 14 files,
# runs pytest, patches syntax errors, and commits cleanly in 1 roundtrip:
# [PTC] Executing batch refactor script (14 operations)... Done.
# [Test] All 38 unit tests passed in 1.4s.
# [Commit] Created branch feat/jwt-rs256 (commit 8a1e2f).

Benchmark Comparison: Terminal-Bench 4 Performance

To validate the impact of Programmatic Tool Calling, Factory evaluated Droid 1.0 on Terminal-Bench 4, the industry standard benchmark measuring end-to-end command-line bug resolution in real Linux environments.