ConnectHubBox unveils Architect for building agent workflows in plain language
Describe a process and Architect maps the steps, decisions and hand-offs, then tests and runs it on your content.
The announcement
Box announced Architect at its BoxWorks 2026 conference in San Francisco on November 5. It is a tool for designing, testing and running AI agent workflows that work on the documents a company already keeps in Box.
An agent does a task rather than just answering a question, for example reading a contract, pulling out the key terms and passing them on. A workflow strings those tasks together into a whole business process, such as onboarding a client or handling an insurance claim. Until now, building one usually needed specialists. With Architect, you describe the goal in plain language and it maps the steps, the decision points and the hand-offs between people and agents.
Design, test, deploy
Every workflow goes through three stages:
- Design: you describe the process and Architect drafts the steps, using procedures and content already stored in Box.
- Test: before going live, teams run the workflow on real documents and awkward edge cases, preview what the agents produce, see where work is routed and add human sign-off at key points.
- Deploy: the workflow launches with the forms, templates and apps it needs. Architect then tracks results and suggests improvements, and admins can make a workflow the company standard.
Your rules apply, and you pick the model
Box's existing permissions, sensitivity labels and retention rules apply at every stage. Teams can also choose a different AI model for each step, using a cheaper one for simple tasks and a stronger one for complex reasoning. Options include models from Anthropic, OpenAI, Google, NVIDIA and IBM.
Box cites contract review at law firms, customer onboarding at banks and claims processing at insurers as examples. Chief executive Aaron Levie argued the real gain is in redesigning whole processes, not just making individuals faster.
Security for agents
Alongside Architect, Box released agent security controls that are available now. Admins can limit which files agents may open based on how sensitive they are, restrict what outside AI tools can do inside Box, and get alerts when agent activity passes a threshold. New guardrails for MCP, the standard that connects assistants such as ChatGPT and Claude to apps, can block those tools from sharing files outside the company.
Threat protection, planned for early next year, targets four risks: bulk copying of files, hijacked login sessions, full account takeovers and mass deletion. It will tie sessions to a specific device and respond in proportion, from restricting downloads to asking the user to verify their identity to ending the session.
Search built with NVIDIA
Box also announced Context Retrieval, an AI search engine built on NVIDIA's Nemotron models and due in early 2027 for Enterprise Advanced customers. It indexes files, including scans, images, audio and video, and hands an agent only the passages it needs instead of whole documents. Box claims 20% better accuracy, 90% less text for the AI to process and answers four times faster. NVIDIA's open-weight Nemotron models, starting with Nemotron 3 Ultra, are available inside Box AI now.
The bigger picture
Most companies are still experimenting. A Cisco survey from March found that 85% of large enterprise customers were trying AI agents, but only 5% used them in day-to-day operations. Competitors are moving the same way: Hyland added tools for managing agents across their lifecycle, and Laserfiche launched agents that build multi-step document tasks from plain language.
Why it matters for your business
Box's design matches what we have learned building agent workflows: describe the process first, test it on real and messy documents, and put a person in the loop where mistakes are costly.
Picking the model per step is how we build in n8n too. A fast, low-cost model handles routine steps like sorting and extraction, and a stronger model is used only where accuracy matters. That keeps running costs down without lowering quality where it counts.
The Cisco figure, 85% experimenting but only 5% in production, is the gap most businesses are stuck in. Tools like Architect help, but the hard part is still choosing the right process and mapping it properly before automating it. Automating a messy process only makes the mess happen faster.
What to do next
- Write down one document-heavy process step by step, including who decides what
- Mark which steps are routine and which need judgement or sign-off
- Test any agent workflow on your messiest real documents before going live
