AI AI Toolkit
AI Newsai-products

Factory 发布 Droid 1.0,以程序化工具调用构建企业级 Agent harness

Factory 研究 / 产品(RSS)2026-10-29T00:00:00.000Z

Key Highlights

Factory released Droid 1.0 (Frontier Harness), positioning it as the "new engine for enterprise coding agents." Its positioning is not another AI model but everything around it — prompts, tools, how tool results are returned, and the run loop. Most distinctive is its model-agnostic design: it supports all frontier open and closed models, lets users bring their own API keys, or has Factory Router automatically pick the best price-performance model mix.

What Happened

Droid is built for complex projects and daily development, meeting enterprise demands for high-quality results on hard tasks, efficient resource use, and strong safety controls. Agents can run continuously for hours or days across hundreds of turns, resending instructions and history each request for consistency. Internally Factory runs an automated loop — new model, evaluate, read failure traces, hypothesize, controlled cross-model experiments, engineer review and merge — so improvements generalize across model families.

Technical Details

Droid 1.0 enables Programmatic Tool Calling (Script mode) by default for a dozen models including GPT-6.1 Sol and Claude Opus 5.5: the agent writes JavaScript to call tools in a sandbox, bringing three advantages — clean context, fewer model requests (15 tool calls drop from 10 requests to 5), and deterministic loops that improve reliability, cutting 15% cost on Claude Opus 5.5. On Terminal-Bench 4, GPT-6.1 Sol scores 3.5 points higher than Codex at 12% lower cost; Claude Opus 5.5 beats Cursor/Claude Code by 3.0 to 7.6 points at up to 24% lower cost.

Versus Competitors

Against Cursor or Claude Code harnesses tuned to their own models, Droid's model-agnostic design plus programmatic tool calling and bounded autonomy (nudge/retry/stop rules at the loop layer, not the model layer) favors enterprise deployment in long-run stability and cross-model efficiency. Prompt-cache hit rate reaches 98% in production, with base prompts shrunk 37%.

Industry Impact and Use Cases

Droid 1.0 ships macOS/Windows desktop apps and a terminal UI, and releases the Droid SDK (@factory/droid-sdk), letting developers launch multi-turn sessions on the same harness, stream progress, and extend custom tools via MCP. For teams that want to scale coding agents inside the enterprise, Factory offers a route that "does not lock you to a single model and emphasizes control and stability."

Practical Notes

The most underrated line in the announcement is the internal improvement loop: Factory treats its own harness as a product it dogfoods against every new model, so gains from one model family carry to others. For enterprise buyers, that means the harness is a moving target that gets better without you changing anything — a different value proposition from a model vendor that only improves when its own weights do.

What to Watch

Two things to watch. First, "model-agnostic" is only as good as the Router's routing quality; if it picks a weak model for a hard subtask to save cost, enterprise users will lose trust fast, so expect fine-grained routing policies to become a configuration battleground. Second, programmatic tool calling shifts complexity into JavaScript the agent writes at runtime — powerful, but it expands the surface for sandbox escape and dependency confusion, so the sandbox's integrity is now part of the security story.

Closing

Droid 1.0 is a bet that the durable value in coding agents is not the model but the harness around it — the part that makes models reliable, affordable, and controllable at enterprise scale. If that bet holds, the category splits into "model companies" and "harness companies," and Factory is staking its claim on the latter.

The Bigger Picture

Droid's arrival reflects where the coding-agent market is heading: away from "which model is smartest" and toward "which system makes models dependable at scale." Enterprises do not primarily lack intelligence; they lack predictable, auditable, cost-bounded execution they can put on a critical path. A model-agnostic harness that quietly improves as new models ship is a bet that the durable moat is the surrounding machinery, not the weights. If that holds, we should expect more "harness companies" and fewer claims that a single model wins outright — and buyers should evaluate vendors on operational maturity, not leaderboard trophies that melt the moment a newer model appears. The companies that win enterprise trust will be the ones that make the boring parts — limits, approvals, resumes after failure — feel effortless, because that is what separates a demo from something a CIO will put on a critical path.