Blog
Box ArchitectBox AI agentsAaron Levie Box Works

What Is Box Architect? How Box's AI Designs Agents

Box Architect uses frontier AI models to design, run and optimize agents and workflows on Box. What it does, why Box built it, and what it means for you.

Portrait of Luis Chavez-Mattos
Edited by Luis Chavez-Mattos, Director, Product··RSS

For IT, Operations and Executives

Three comic panels: a grinning analyst, her snapping large building blocks into a chain, and a close-up of two blocks clicking together with one orange block.

What is Box Architect?

Box Architect is an autonomous system from Box for designing, building, running and eventually optimizing AI agents, workflows and apps that sit on top of Box. Instead of a team sketching a contract management or invoice processing solution on a whiteboard, you describe the goal and a frontier AI model designs the full solution from Box's own building blocks. CEO Aaron Levie announced it at Box Works. The optimization part, where it learns from failures, is described as coming later.

TL;DR

  • Box Architect designs the solution for you. You name a process (contract management, invoice processing, supply chain document review) and an AI model drafts the whole agent and workflow setup.
  • It builds from pieces Box already has, including its workflow tool, its no-code app builder and its document management system.
  • Box says the idea came from a prototype. An engineer's early test produced in about 30 seconds what customers spend roughly two days designing.
  • Optimization is the next step, not day one. Levie said that "in months" it should review failures and improve the process based on ground truth.
  • It follows a pattern. Decagon and Sierra have agents that design customer support agents. Architect applies the same idea to any business process.
  • Box also stays open. Customers can use Box as a "headless" content layer with other tools instead of using Architect.

What does Box Architect actually do?

Box is best known as a place where companies store and govern documents: contracts, invoices, supply chain paperwork, marketing plans. Most of that is unstructured data, meaning files rather than neat database rows.

Architect sits on top of that. Levie's examples of what an enterprise might ask it to design:

  • A contract management solution
  • A supply chain document review solution
  • An invoice processing solution
  • Customer onboarding
  • Loan processing
  • Client intake

Architect uses whichever frontier AI model is leading at the time to design the complete solution. It assembles that solution from building blocks inside Box. Levie named three that Box already had: a workflow tool, a no-code app building tool and the core document management system.

The alternative, in his words, is to build a lot of that software yourself or to "cobble together three or five different tools" to get the same result.

Why did Box build it?

The origin story is a surprise at a whiteboard. Levie said an engineer built an experiment six to nine months before the announcement. It asked the latest models to design a full Box solution, the kind of app a customer would build. The agent did in 30 seconds what customers spend about two days on: thinking it through, drawing it up, and mapping the whole architecture.

Two things followed from that.

The models kept improving. Levie said they got much better at using tools, operating computers and understanding business processes. He pointed to benchmarks such as GDPval, which test performance on enterprise-style work. Box decided it would not have launched Architect until the models were ready, and this summer's progress convinced the company they were.

The bigger idea is about process, not speed. Levie's argument is that companies shouldn't automate today's work exactly as it looks and drop agents into it. An AI may be able to suggest a better way to run the process in the first place. Box expects that, at some point, models will design workflows better than people do.

How is Box Architect different from Decagon and Sierra?

It isn't a rival so much as the same idea in a wider setting. Levie compared Architect to Decagon and Sierra, which offer agentic systems that help a customer support operations person design the customer-facing support agents. A third vendor does something similar. This approach has caught on most in customer support automation.

The structure is the same in each case: an agent helps a human operator design other agents. The difference is scope.

Support-agent designersBox Architect
Who uses itCustomer support operations teamsAnyone automating a business process
What it designsCustomer support agentsAgents, workflows and apps on Box
Where it worksSupport toolsBox content and building blocks

You tell Architect to design the next agentic workflow for a process, and it designs the whole thing.

How will it optimize workflows over time?

Levie was careful about timing. Design comes first. "Eventually" and "in months" were his words for optimization. The plan is that Architect looks at failures in a running workflow and improves it based on ground truth, so it becomes an autonomous helper for the workflows you've automated.

Treat that as a stated direction, not a shipped feature. The source gives no details on availability, pricing or how the improvement loop works, so ask Box before you plan around it.

Why does Box keep throwing things away?

Levie described himself as "in the upper limit" of what the AI community calls bitter lesson pilled. The bitter lesson is the observation that general methods that scale with computing power tend to beat clever hand-built tricks. In practice, Box tries to ride the improvement curve of frontier models. It builds little that a better model would make obsolete, except temporary scaffolding, meaning code that exists only because today's models can't yet do something.

Asked how often that scaffolding gets discarded, his answer was "all the time." He said Box has had unreleased capabilities that a new model made useless before they shipped.

What Box treats as lasting is the structure a safe AI deployment needs even with far smarter models:

  • Data governance
  • Strong security
  • Fast, high-throughput infrastructure agents can use
  • Easy-to-use schemas and tools, so an agent can create objects in your system

Levie compared this to the "API first" shift of 15 years ago. Many scenarios will involve nobody ever seeing a software vendor's interface, so the product has to deliver value without one.

Why does the context layer matter for agents?

Levie cited Jensen Huang's rough estimate of 10 to 100 times more agents than people. Someone has to decide what data those agents can reach: CRM data, ERP data, lakehouse data, files. He argued you can't toss everything into one giant pile. Access controls, governance, metadata and sandboxing are what tell an agent where to look and what it may touch.

That is Box's pitch. Architect designs the agents, and Box's content layer controls what they can see.

What does this mean for you?

If you work in operations, legal, finance or procurement: Processes like invoice handling, contract review and client intake are the ones named. Expect designs that start from a goal rather than a blank diagram. Your value shifts toward knowing the process, spotting what the AI got wrong and defining what "correct" looks like.

If you run IT or buy software: Ask three questions. What data can the agents access? How are failures reviewed? What does it cost? Levie himself noted that you can't predict token spend in advance, which will worry finance teams.

If you already store content in Box: Architect is the shortest path from "we have the documents" to "we have an automated process." If you use other systems, Box says it remains open and will work as a headless layer.

What to do now: Pick one document-heavy process. Write down its steps, its exceptions and who approves what. That description is what any AI designer, Architect or otherwise, needs, and writing it will show you where your process is unclear.

Can you build a small version of this yourself?

Not everything needs an enterprise platform. Many teams only need a place to track a process, such as an invoice approval queue, a contract renewal tracker or a client intake form. Remy builds and deploys full-stack business apps from a plain-English description, with a database, sign-in and a live URL included. You could describe your intake form, the approval steps and who can see what, then get a working app you own instead of another spreadsheet or subscription. It doesn't replace a content system like Box, but it suits the small internal tool next to it.

Frequently Asked Questions

4 questions

Is Box Architect available now?

Box announced it at Box Works. The source doesn't cover pricing, availability tiers or rollout dates. Levie described the optimization feature as coming in months, so check with Box for current details.

Does Box Architect replace people who design workflows?

Levie's view is that models will eventually design workflows better than people do, and that AI should help redesign the process itself. The source doesn't say what happens to specific jobs. Practically, people still define goals, review the design and judge whether results are right.

Which AI model does Box Architect use?

Box says it uses whichever frontier model is leading at a given time. The source doesn't give a fixed model, which fits Box's approach of not betting on one model.

Do I have to use Architect to use Box with AI?

No. Levie said Box is "relatively indifferent" to whether you use Architect or use Box as a headless solution with other technology. Box wants both to be convenient, and also offers an end-to-end option for customers who lack the time to assemble the pieces.

How we source, write, and correct our articles. Editorial standards

Related articles

Straight to your inbox.
New pieces as they publish. No noise.