Playbooks · Coding tasks
Factory Droid 1.0 and the harness-efficiency claim
Written November 4, 2026
Factory's harness cut prompt tokens 37% and reports fewer model calls than direct tool use. All figures are the vendor's; verify first.
What changed
Factory introduced Droid 1.0 on 29 October 2026 as the largest update yet to its coding-agent harness — everything around the model: the prompts, the tools, how tool results return, and the loop that runs it all. Droid is model-independent: it tunes its prompts, tool descriptions, and tool-calling mode for each model it supports, including models used with a user's own API keys. The headline change is programmatic tool calling, now the default, implemented as a Script tool. Script mode is on by default for GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Astra, GPT-6 Sol, GPT-5.6 Sol, Claude Fable 5.1, Claude Fable 5, Claude Opus 5.5, Claude Opus 5, and Kimi K3.
How Script mode works
Instead of calling tools one at a time and waiting for each result to return to the model, the model writes a JavaScript program that runs in a sandbox with no file, network, or system access of its own. A single Script can make hundreds of tool calls in one model request, so the instructions and history are sent once rather than re-sent every few calls. Raw tool results stay inside the program as variables; only printed output, capped in size, returns to the model, and the full results are logged where the model can search them later. Calls that do not depend on each other run concurrently. Factory's reasoning is that checks and loops then run deterministically in code, so a planned case cannot be quietly dropped mid-task.
What Factory reports
Every figure in this section is Factory's own, from its own sessions, telemetry, or benchmark runs. None of it has been independently reproduced, and the direct-call comparisons in the two worked examples are Factory's estimates, not measured runs.
- A smaller first request. Between Droid 0.170.0 and 0.250.0, the first request shrank from 20,887 input tokens to 13,055, a 37% reduction. Droid now sends 2 tool definitions up front instead of 11, loading rarely used tools on demand. Newly loaded tools are appended to the end of the request so the cached prefix is left untouched. Factory reports that the typical production request reads 98% of its input from cache, and that the token-weighted share of input read from cache was 91.5% over 20.4 million production requests in the seven days to 22 October.
- Fewer requests on two real sessions. A Linear cleanup moved 151 issues between queues and verified each move, 312 tool calls in all. The Script took 2 requests and about 61,000 tokens; Factory estimates direct calls would have taken 36 requests and about 3.1 million input tokens at 10 calls per request. A CI sweep checked 5,384 desktop jobs across 547 runs, 550 tool calls, in 2 requests instead of an estimated 59 to 187, with under 0.1% of the tool results reaching the model's context.
- Lower cost on one comparison. Compared with direct tool calling, Script mode cuts Droid's cost by 15% on Claude Opus 5.5, with the savings concentrated in the 12.5% of Scripts that make six or more tool calls.
- Terminal-Bench 4 head-to-heads. With GPT-6.1 Sol, Factory reports Droid scoring 3.5 points higher than Codex at 12% lower cost, and 6.6 to 12.6 points higher than OpenCode and Devin. With Claude Opus 5.5, it reports Droid scoring 3.0 to 7.6 points higher than Cursor, OpenCode, Devin, and Claude Code, at up to 24% lower cost. Each harness ran three trials per task between 23 and 28 October 2026, at high reasoning effort, with cost and time measured per passed task at list prices. The comparison counts only tasks that both harnesses passed in all three trials. Two methodology details matter: tasks where Devin stopped to ask the user a question were counted as passed if Codex passed them, and Cursor's token counts in two trials were estimated from transcripts because Cursor undercounts tokens when it reconnects mid-task.
Bounded autonomy
Long runs get stuck, repeat work, or drop the task, and Droid puts the recovery rules in the loop rather than in the model, so they hold for every supported provider. Three identical tool calls get a nudge and six stop the loop; three turns in a row with no visible output also stop it, while repeats that return new results, such as polling a build, can continue. When a long conversation is summarized, the compaction summaries now have room for 16,000 tokens, up from 8,000, because Factory estimates a third or more were filling the old limit.
How the harness is tuned
Factory's internal Agent Factory loop is part of the claim: agents study benchmark failures alongside friction from real sessions, propose changes, and test each one in a controlled experiment on every model family it affects. Engineers review the evidence and decide what ships. That loop produced Script mode and the per-model choice of whether to use it, and it keeps producing most of Droid's changes. A builder reading the efficiency figures should read them as the output of that loop — tuned against the same benchmarks used to report the gains — not as a neutral measurement.
What to do
- Do not budget from the percentages. A 37% smaller first request is not a 37% smaller bill: most of a long session's cost is history and tool results, and how much of that is cached depends on the provider and the session. Run Droid on your own workload and compare cost per finished task against the harness you use now.
- Check Script mode on your model. The default list covers ten models. If yours is not on it, the Script-mode figures do not apply to your setup, and the per-model tuning may differ.
- Judge the worked examples by their comparisons. The Linear and CI examples compare a real Script session against an estimated direct-call session. They show why batching helps; they do not measure how much it helps on your tasks.
- Watch for repetition on long runs. The nudge-and-stop rules are the safety net for loops that repeat themselves. If an agent repeats the same call without new results, that is the behaviour the rules exist to catch — stop it rather than letting it run.
Limits
- Factory Droid is proprietary. There is no independent run behind any figure on this page, and no way for a reader to reproduce the Terminal-Bench comparison without Factory's setup.
- The benchmark comparison covers two models at high reasoning effort on one benchmark. It says nothing about other models, lower reasoning settings, or tasks outside Terminal-Bench 4.
- The prompt-token cut, cache shares, and latency figures come from Factory's production telemetry and controlled tests, not from an independent measurement. Treat them as the vendor's account of its own system.
Related playbooks
- Factory Droid 1.12.0: the November build
- AIDE2's self-improvement claim vs Epoch's evaluation
- VS Code 1.146 retires the Local harness
- Qwen Code 1.0 folds in the Qoder harness
Sources
Back to: Coding playbooks hub · Also: Cut coding-agent costs with TokenEconomizer · BrokenML: a benchmark of broken machine-learning jobs