Source: raw/Box_CEO_Explains_Why_AI_Makes_Software_Features_Obsolete.md

Creator: Matthew Berman (guest: Aaron Levie, CEO of Box, recorded at BoxWorks) · URL: https://www.youtube.com/watch?v=kMdHCyAnNHA · Duration: not stated in the transcript (about 7,000 words) · Platform: YouTube (auto-captions, fetched 2026-11-17)

A CEO describing his own company’s product rules for a period when each model release can make a feature pointless. Levie says Box has shelved unreleased features because a new model overtook them, and that he builds only what would still be needed if models were a thousand times better. He also covers two same-day Box announcements, how he expects token spend and pricing to settle, and why cheap open-weight models raise spending on frontier models. The recording date is not stated; Berman congratulates him on an all-time-high day for the stock.

Key Takeaways

  • “Upper limit bitter lesson pilled.” Box wants to “ride out the curve of intelligence” and avoid building anything better models would make obsolete, apart from temporary scaffolding.
  • Features have been thrown away before release. “We couldn’t release it soon enough before a new model just made it totally useless.” His example: three months on a layer that became unnecessary once models got good at tool use. “You have to be very unsentimental.”
  • What survives better models. His test is to assume superintelligence and ask what an organisation still needs to deploy it safely: data governance, strong security, fast high-throughput infrastructure agents can call, and simple schemas and tools for creating objects in the system.
  • Assume nobody sees your interface. “There’s going to be a very long list of scenarios where nobody will ever see your interface,” so the product must be valuable headless. He compares it to going API-first 15 years ago.
  • Two layers of software. Many agents on top; underneath, systems of record that control access, governance and storage, and still keep a user interface because people have to put data in. He cites a figure of 10 to 100 times more agents than people.
  • Context cannot be a blob. Agents need access controls, governance and metadata so they know where to route a query; dropping everything into one store does not work.
  • On token spend, target beats volume. He is “not a fan” of token-maxing: “the correct application of tokens matters probably the most”. He does think compute pointed at worthy problems becomes a competitive advantage, and names the risk that capital then decides who wins.

What Box announced

  • Box Architect. An autonomous system that designs, builds and runs agents, workflows and apps on Box, with optimisation from observed failures to come “in months”. A customer asks for, say, a contract management or invoice processing solution, and a frontier model (he names Astra, Opus and Fable as today’s candidates) assembles it from Box’s building blocks.
    • Origin: six to nine months ago an engineer had a model design a full Box app. “The agent just did in 30 seconds what our customers spend two days thinking about, designing on a whiteboard.”
    • Why now: Levie says Box would not have launched it before this summer’s gains in tool use and computer use.
    • The idea behind it: do not automate today’s process as it stands and drop agents in; let the model propose a better process.
    • Comparables he names: Decagon, Sierra and Fin, whose agents help a support-operations person design the customer-facing agent.
  • Context retrieval engine, built with Nvidia. A pre-indexing layer using Nvidia’s embedding model to rank results and return only the relevant chunks, so agents spend fewer tokens on unstructured data. Available invisibly or through an API that returns chunks. Rolling out “over the coming months” in the Enterprise Advanced edition. Berman calls the model “NeMo Tron retrieve”; that spelling comes from captions.
  • Own models. Levie thinks post-training for some use cases is “almost inevitable” but Box is still riding base-model gains. He names post-training vendors he wants to work with: Trajectory, Prime Intellect, Applied Compute, Fireworks AI and Ngram (names as captioned).

Token budgets and pricing

  • His three-year picture: every piece of information produced gets reviewed by “1, 3, 5, 10 agents”. Examples: an always-improving profile for every sales account; agents investigating each odd login and each code push; agents testing other agents’ code; agents extracting data from every supply-chain document.
  • How far along: “probably like a hundredth of the way” to agents being attached to our digital interactions.
  • Pricing: keep a per-seat component while people work with the data, for predictability, and add consumption pricing because agent work varies in ways no one can forecast.
  • Budgeting: he expects token spend in three to five years to be an ordinary line item, raised or lowered like a marketing budget. For now, “you can’t really predict in advance” what a problem will cost in tokens, and Box’s contribution is cheaper retrieval.

Open weights

  • Offer anything safe, default by evals. Box offers open-weight and closed models and uses its evals to choose defaults. His ideal is a routing layer that picks the model for each query at the lowest cost for the accuracy required.
  • Cheap open models increase frontier spend. A closed frontier model plans and orchestrates; open or specialised post-trained models do the bulk work, such as reading every document or checking every transaction for fraud. The bulk tier makes the whole job affordable, so the frontier tier gets used where it otherwise would not.
  • Both grow together. His analogy, which he calls poor: startups after 2005 ran on open-source databases while Oracle still became a $300 billion company.
  • Who can fund open weights: whoever profits from more inference. He names Meta, Google and Nvidia.
  • One generation back is enough. If open weights trail the closed frontier by three to six months, they take over each use case the frontier had solved six months earlier. He says his “80% of IT ticketing tasks” figure is made up to illustrate the point.
  • Open weights as a check on labs. Asked whether labs gain too much power as gatekeepers of tokens, he calls open weights the “escape valve”.

Try It

  1. Run his test on your own stack. For each prompt chain, wrapper or workaround you maintain, ask whether a much better model would still need it. Governance, access control, clean data and tool schemas pass; glue for a current model weakness is temporary.
  2. Split agent work into two tiers. Give planning to a frontier model and bulk reading or classification to a cheaper model, then compare total cost with running everything on one model.
  3. Budget tokens by problem, not by headcount. Pick the few tasks where more compute changes the outcome and fund those, as he does when he rejects token-maxing.
  4. If you sell a service, check its value without your interface: what would a client’s agent call, and what would it get back?
  5. Before automating a client process as it stands, ask a frontier model to propose a better process first.

Open Questions

  • Dates and prices. No recording date, no Box Architect price and no general-availability date are given.
  • Meta’s open-weight model. Levie refers to it as “Muse, Spark, Open or something like that” and is unsure of the name; Berman places it at roughly the closed frontier of March 2026. Neither names it exactly.
  • The 10 to 100 times agent figure is attributed to Jensen Huang “on stage” without a source.
  • Gavin Baker’s chart of token share moving to open models is mentioned by Berman and not shown.