Droid 1.0: Script-Based Tool Calling Cuts Agent Costs by 50x on Batch Tasks

Forum topic · QianXun · 2026-10-31

Summary

Factory released Droid 1.0 on October 29, an agent harness that replaces per-step tool calling with programmatic execution: instead of issuing hundreds of individual tool calls, the model writes a JavaScript script that runs in a sandbox and returns only a summary. In a benchmark task involving moving 151 tickets via 312 tool calls, the traditional loop consumed roughly 36 model requests and 3.1 million input tokens, while Droid completed it with 2 requests and about 61,000 tokens. Factory's self-reported benchmarks claim Droid outperforms Codex by 3.5 points at 12% lower cost with GPT-6.1 Sol, and beats Claude Code, Cursor, OpenCode, and Devin by 3.0–7.6 points with Claude Opus 5.5, up to 24% cheaper. Startup instructions were cut 37% to 13,055 tokens by loading 2 tool definitions up front instead of 11, and the company says 98% of production input tokens come from provider caches. The approach builds on the 2024 CodeAct paper. All figures are vendor-tested without third-party replication.

The problem: paying to resend context on every tool call

In a coding agent, the model only predicts text; what actually lets it read files and run commands is the outer layer — the "harness" (instruction templates, tool definitions, and the execution loop). Codex and Claude Code are harnesses. Under the conventional loop, every tool call requires a full round trip in which the instructions and history are resent. A batch job moving 151 tickets required 312 tool calls, roughly 36 model requests, and about 3.1 million input tokens.

Droid 1.0: the model writes a script instead

Droid 1.0 (released October 29 by Factory) changes the loop: the model writes a small JavaScript program, the program calls the registered tool interfaces itself, and only a result summary is returned to the model.

The script executes in an isolated sandbox with no access to real files or the network — only registered tool interfaces.
The 151-ticket migration now takes 2 model requests and ~61,000 tokens — a ~50x reduction.
In another task, the model verified 5,384 test jobs across 547 runs while only ever seeing a summary flagging 41 failures.

The idea of programmatic tool calling is not original to Factory; the company credits the 2024 CodeAct paper (agents interacting with environments via code rather than single instructions). Droid's contribution is making it the product-level default.

Benchmark results (Factory's own tests)

| Comparison | Droid advantage | Cost difference | |---|---|---| | Droid + GPT-6.1 Sol vs Codex | +3.5 points (Terminal-Bench 4) | 12% lower | | Droid + Opus 5.5 vs Claude Code | +3.0 to +7.6 points vs Claude Code, Cursor, OpenCode, Devin | up to 24% lower | | Time per solved task | 26% faster than Claude Code | 17% fewer tokens | | Speed on GPT-6.1 Sol | 10% slower | 12% fewer tokens |

Notable details:

Switching to script-based calls alone cut 15% of cost on Opus 5.5.
On Anthropic models Droid is both faster and cheaper; on OpenAI models it saves tokens but is 10% slower — harness–model fit varies by vendor.
Devin's habit of pausing to ask the user questions was handled by crediting Devin for tasks Codex solved where Devin paused (4.7 questions per trial on average). This adjustment affects accuracy only, not cost or time.

Other optimizations

Startup instructions compressed from 20,887 to 13,055 tokens (–37%) by pre-loading only 2 of 11 tool definitions, loading the rest on demand.
Factory claims 98% of production input tokens are read from provider caches, which are priced well below full rate.

Takeaway

The underlying bet is model neutrality: a single harness tuned for all models (Anthropic, OpenAI, open-weight like Kimi K3) rather than vendor-tuned shells. The architectural signal is the key point — compute for batch operations should go to doing the work, not resending context. All numbers are vendor self-tests with no third-party replication.

---

Sources and caveats: Factory's official announcement "Droid 1.0: Building a Frontier Harness" (October 29) and item-by-item verification by AI Insiders. All benchmark figures are vendor-tested; the Devin scoring adjustment (4.7 questions per trial) comes from Factory's own methodology notes; the "98% cache read" figure is a company claim; the CodeAct paper attribution is per Factory.