Box’s CEO throws away features on purpose

Box just shipped a tool that designs your business workflows for you. The bigger surprise is how the team decides what to build and what to throw away. I just watched an interview where the creator, Matthew Berman, sat down with Box CEO Aaron Levie at BoxWorks, and it changed how I think about building on top of AI.

Here’s the breakdown.

🚀 What’s new: Box Architect

Box Architect is an autonomous system that designs, builds, runs and, before long, optimizes your agents, workflows and apps on Box. You just ask for what you want, for example “design me a contract management solution” or “build an invoice processing flow.” It pulls the building blocks from Box and puts together the whole solution using whichever frontier model is leading at the time.

The origin story is wild. Six to nine months ago, a Box engineer ran a quick test. Levie said the agent “did in 30 seconds what our customers spend two days” planning on a whiteboard. Once the summer’s model upgrades landed, the team decided AI should help redesign business processes, not just automate the ones you already have.

Box also announced a new context retrieval engine built with NVIDIA. It pre-indexes your unstructured data so agents only grab the chunks they need. That means fewer wasted tokens and faster answers. It’s rolling out to the Enterprise Advanced edition over the coming months.

🔄 The twist: build things you can throw away

This is the part that stuck with me. Levie calls himself “upper limit bitter lesson pilled.” The “bitter lesson” is the AI idea that general methods riding on more compute end up beating hand-built cleverness.

In practice, Box has built features it never released because a new model made them useless first. Three months of work, gone overnight once the models got good at tool use. His advice is to be totally unsentimental about it.

So what’s worth building? Only the things you’ll still need when models are a thousand times smarter:

  • Data governance and access controls
  • Strong security
  • Fast, high-throughput infrastructure for agents
  • Simple schemas that agents can plug into

He compares it to Amazon’s old API-first push. Assume plenty of users will never see your interface, and make sure you still deliver value without one.

🛠️ A mini-workflow based on his thinking

  1. Before you build, ask: “Would a much smarter model make this pointless?” If the answer is yes, treat it as temporary scaffolding.
  2. Put real effort into your context layer. That means permissions, metadata and governance, not one giant blob of data.
  3. Let AI propose a better process before you automate the old one.
  4. Mix your models. Use a closed frontier model as the planner, and use cheaper open-weights models for high-volume work like checking every document or every transaction for fraud.
  5. Treat tokens like a normal budget line, the same way you’d plan spending on marketing.

💡 Pro tips from the conversation

  • Throwing the most tokens at a problem isn’t the goal. Where you point the compute matters more than how much you use.
  • Cheap open-weights models can actually raise your frontier spend, because suddenly bigger projects become affordable.
  • Open models that run three to six months behind the frontier still pick up a huge share of real work, like most IT ticketing.
  • Expect pricing that mixes predictable seats with usage-based charges for agent work.

Levie’s big prediction: within three years, everything you produce will be reviewed by somewhere between one and ten agents. He thinks we’re about “a hundredth of the way” there.

👉 Watch the full interview for his take on open weights, AI labs as gatekeepers, and the “bitter lesson” memo he says he plans to publish.

Scroll to Top