Factory introduced OpenCompute, a self-hosted orchestrator for agent sandboxes that runs on customer-controlled infrastructure, now in private preview. It brings on-demand provisioning and lifecycle management to customer environments, integrated with Factory's Environments, Automations, and Remote Control features. This allows enterprises to manage fleets of agent sandboxes on their choice of runtime (Kubernetes, microVMs, etc.) without handing control of their infrastructure to Factory.

Factory is introducing Environments, now generally available, which are reusable sandbox definitions for Droids composed of repositories, tools, and dependencies. Environments are pre-built into portable images so each sandbox is ready before a Droid arrives, enabling instant onboarding and parallel task orchestration. The release also includes integration with Automations and Agent Readiness frameworks for improved fleet management.

Factory launched a redesigned Analytics feature providing engineering leaders with full transparency into consumption, model efficiency, and adoption across every session. The new Analytics includes Consumption views broken down by model and user, Model Router savings reporting (showing 63% cost reduction), and Adoption tracking with daily/monthly active users. An Analytics API is also available for enterprise BI tool integration.

Automations are now generally available in Factory. Users can describe a recurring workflow, choose a schedule or event trigger, and let Droid run it automatically.

Factory's Legacy-Bench benchmark joined Fireworks' Specialized Intelligence Index (SII), enabling comparison of open, closed, and specialized AI models on legacy software engineering tasks. The benchmark covers debugging, implementation, and migration across COBOL, Java 7, BASIC, C89, Fortran, and Assembly. Ten representative tasks are now publicly available on GitHub in the Harbor format.

Factory introduced flexible deployment options with Factory Private, allowing organizations to run Factory in their own VPCs, on-premises environments, or air-gapped networks.

Factory is now listed on Anthropic's Claude Marketplace, allowing enterprise customers to apply their existing Anthropic committed spend toward Factory without a separate buying cycle. The integration supports connected deployment models where customers can use Claude through their approved Anthropic setup while keeping execution close to their own infrastructure.

Factory is expanding into Asia by opening an office in Tokyo, Japan, to meet growing regional customer demand. Seiji Sasaki, formerly Site Lead at OpenAI Japan and VP/GM at Slack Japan & Korea, has been appointed as President & GM of Japan. Factory is actively hiring across Sales, Marketing, Support, and Partnerships in the region.

Factory introduced a multi-role agent system where one role builds an executable standard of completion before implementation begins, and the implementation is held to that standard. This approach significantly improved behavioral parity on large software recreation tasks compared to single-agent runs, e.g. taking 7-Zip from 54% to 95% and DuckDB from 34% to 80% parity. The system was evaluated across 24 ProgramBench tasks and three models.

Factory launched Connectors, giving Droid managed access to apps like Linear, Slack, Sentry, Notion, Salesforce, GitHub, Datadog, Figma, Amplitude, Stripe, and more. Personas were introduced so each role (Engineering, Product, Design, Finance, Marketing, Sales, Operations) is guided to the right Connectors during onboarding. Enterprise Owners and Managers can control which Connectors are available and recommend them per persona, while Autonomy Levels govern when Droid requires confirmation for actions.

Factory launched Factory Router, a model routing system built into the agent harness rather than at the gateway layer. The router has been in production for over two months, cutting aggregate inference costs 58% compared to using frontier models for every call, with median routed sessions saving 76%. Routed sessions matched frontier-pinned sessions across eight production quality measures while reducing median turn latency from 81 seconds to 49 seconds.

Factory announced the Factory Partner Network (FPN), backed by a $100M commitment to partnerships across training, deployment, solutions, and marketing. The program enables consulting & SI firms, CSPs, ISVs, and AI labs to plan, deploy, and operate Factory's software factory for enterprise customers. Factory will cover costs of qualifying deployment activities and provide partners with exclusive access to sales, engineering, and marketing resources.

Factory has opened applications for the Factory Guild, a curated community program for students, independent builders, and developers who want to build with Droids. Members earn points and advance through ranked tiers (Apprentice, Craftsman, Artisan) by contributing projects, content, events, and community support. Top contributors may be considered for roles at Factory.

Factory launched Agent Effectiveness in Analytics, connecting Droid sessions to cycle time, work intent, and shipped artifacts. It provides three views—Throughput, Output, and Attribution—so engineering leaders can measure AI ROI by linking Factory usage to delivery outcomes. The feature integrates with Jira, Linear, GitHub, and GitLab and requires enabling the Advanced Analytics enterprise control.

Factory now supports autonomous software engineering on NVIDIA DGX Spark using NVIDIA Nemotron 3.5 Lightning, enabling high-security and air-gapped environments to run the software factory locally. Teams in regulated industries such as banking, government, and critical infrastructure can connect Factory to an on-premises deployment of Nemotron 3.5 Lightning or other approved open-source models. This integration supports comprehensive engineering tasks including issue investigation, code modification, test execution, and change preparation without data leaving the controlled environment.

Droid Shield 2.0 introduces two fine-tuned models to strengthen deterministic secret scanning. Factory published the LoRA adapter weights, tokenizer, configuration, training prompt, and calibration guidance on Hugging Face so other teams can run, inspect, and build on it.

Factory introduced an AI-native enterprise organizational model that allows large engineering organizations to create nested organization hierarchies under a single enterprise account, each with isolated membership, integrations, model access policies, usage limits, and data residency settings. The model supports BYOK (Bring Your Own Key) isolation per organization, enabling sensitive workloads to run on customer-managed keys while other groups use Factory-managed inference. Enterprise admins gain centralized visibility and governance via a dedicated Enterprise Admin Console.