Latent Space: The AI Engineer Podcast

Episode: "Every Agent Needs a Box — Aaron Levie, Box"

Date: March 5, 2026

Guests:

• Aaron Levie (CEO, Box)
• Host: Latent.Space team
• Jeff Hoover (Chroma CEO, guest host)

Episode Overview

This episode explores the rapidly evolving role of AI agents in enterprises and the implication for infrastructure, data governance, and organizational change. Aaron Levie, CEO of Box, discusses how the rise of agents—AI-powered autonomous or semi-autonomous systems—demands a rethinking of file storage, access control, and collaboration paradigms. The hosts and Levie examine context engineering, knowledge work transformation, and the looming future where "every agent needs a box" to operate safely and effectively.

Key Discussion Points & Insights

1. The Enterprise Agent Wave: "Every Agent Needs a Box"

• Main Theme: The shift toward AI agents in enterprise, where agents need secure, governed access to content—a modern evolution from user-focused file systems.
• Levie’s Central Analogy:

  "Every agent needs a box." (Aaron Levie, [04:15])

• Implication: Box's core value proposition becomes foundational as the "operating system" for agents, governing their access, permissions, and secure collaboration.

2. From Human-First to Agent-First Workflows

• Work Reconfiguration:
  • The nature of software work, especially code, has already radically changed due to AI agents.
  • Other knowledge work (legal, sales, HR, finance) lags because of more complex, analog, and fragmented contexts.

"You don't write code. You talk to an agent and it goes and does it for you, and you maybe at best, review it. ... We are changing our work to make the agents effective. In that model, the agent didn't really adapt to how we work. We basically adapted to how the agent works." (Aaron Levie, [00:00])

• Future Outlook:
  • The rest of the economy must now re-engineer workflows to align with how agents can operate, which will take multiple years and give early adopters a compounding advantage.

3. Security, Identity, and Data Governance for Agents

• Scaling the Number of Agents:
  • Levie predicts “an order of magnitude more agents than people” ([04:30]).
  • This requires new systems for security, permissions, and agent authentication/identity.
• Identity Challenges:
  • Traditional notions of user accounts don’t map neatly to agents.
  • Creators of agents may be liable for their actions, and agent privacy is not analogous to employee privacy.
  • Overlapping access and liability become hairy as agents act both autonomously and collaboratively.

“You can't just be like, oh, I'll just create a bunch of accounts and then I'll kind of work with that agent ... you need oversight of that.” (Aaron Levie, [07:19])

• Security Incidents:
  • There are inevitable “crazy security incidents” as prompt injection and permission exploits emerge ([05:00]).

4. Context Engineering: The New Bottleneck

• Context Rot:
  • The challenge of maintaining, surfacing, and updating relevant information out of massive document stores.
  • Infinite context windows are a fantasy, so smart search, ranking, and context pruning are the real focus.

"How do I bridge the 50 million pages of information with the couple hundred [tokens] that I get to work with in that token window? ... that's why so much work is just search systems." (Levie, [22:08])

• Evaluating Agent Judgment:
  • When should agents give up on fetching info? How do you know if an agent “doesn’t know what it doesn’t know”?
  • Challenge: Agents often fail gracefully—returning partial answers or proceeding despite insufficient context ([25:24]).

5. Knowledge Work vs. Code: Mess, Structure, and Slop

• Coding is Clean(er):
  • Codebases have documentation, structure, and shared context—ideal for agentic workflows.
• Knowledge Work Complexity:
  • Enterprises are “messy, poorly named and duplic[ated], outdated shit” (Levie, [18:33]), making agentic enablement far more difficult.
  • Documentation rot and non-digitized knowledge limit agent usefulness.
  • Tacit knowledge and apprenticeship (“onboarding ramp-up”) are hard to capture for agents.

“Most companies are practically apprenticeships ... all that tacit knowledge is not written down, but it would have to be if you wanted to like give it to an agent.” ([43:56])

6. Evaluating Agents and Team Structure

• Eval Culture:
  • Box runs industry-specific, private evals to test agents across legal, healthcare, finance, and government documents (see [32:39]).
  • Huge advances in model performance across generations.
• Team Organization:
  • Box operates an “internal startup” focused on agents, with core AI, search, and infra teams supporting.
  • Multilayered development, "layers and layers of concentric circles" ([35:39]) around the agent core.

7. Read vs. Write Workflows: The Next Agent Opportunity

• Agent as Creator:
  • Beyond reading/finding content, agents will author, edit, and store all types of files, from docs to presentations.
• Challenges:
  • Write tasks (content generation) are easier technically, but formatting (like PowerPoint) is still a struggle for models.

“Building a beautiful PowerPoint presentation is still a hard problem for these models ... the end user instantly sees it." ([40:32])

8. Knowledge Graphs, Wikis, and “Company as File System”

• Debate: Structured Knowledge vs. Simplicity
  • Some advocate for highly normative “knowledge graph” structures; others for markdown files and wikis.
  • Levie: Not religious about graphs, Box can feed data into any paradigm. Most companies (and users) default to collaborative workspaces, not formal graphs.

“I certainly like the vision of most knowledge graph... Just like it's 2026, we haven't seen it yet... play out as ...the main thing.” (Levie, [47:03])

9. Founder Reflections & The Role of Delegation

• Delegation vs. Existential Work:
  • Levie delegates most of Box’s operations, focusing personally on existential, future-oriented work (like agents/AI).

“If we don’t do this right, we’re out of business. ... this is like you can just see how the AI tsunami could wipe you out if you make just 2, 3, 4, 5 wrong decisions in this space.” ([50:48])

• Maintaining Hands-on Ownership:
  • Important to not entirely step away, or entropy will take over.
  • Continuous check-ins and deep involvement in core innovation initiatives.

10. On Building in Public & the Aaron Levie Production Function

• Content and Communication:
  • Levie is unusually public and prolific with mini-essays and ideas on social platforms; sees this as a natural byproduct of high feedback loops between operating “in the trenches,” reflecting, and sharing.
• Cultural Note:
  • “Building in public ... is just a natural thing. ... I just go through the day. We deal with interesting problems. I tweet about them. I get information back in the Process. ... My job is to try and connect all these things together and, and make it useful.” ([59:05])

11. AI in Film, Creativity & Democratization

• Levie’s Film Background:
  • Got the idea for Box while working at Paramount, seeing the pain of sharing production files.
• AI in Media:
  • Optimistic about democratizing creativity but wary of generative entertainment "slop."
  • Sees promise for educational, niche, and high-quality children’s content.

“I think that ability to just. You get to be Spielberg, you know, is ... completely amazing and democratizing that is incredible. ... I still like the idea of like this is a camera and a person and a person that says action. And let's hopefully surround AI around that.” ([63:03])

12. Media, DevRel, and the Next TechCrunch

• Every Company = Media Company:
  • Companies must become media/DevRel organizations to win developer and agent "attention" and drive adoption.
• DevRel Talent Crunch:
  • Demand is outpacing supply (“If you could produce a freaking factory of DevRel people, there’s just like unlimited jobs” – [69:56]).
• Advice for Latent Space:
  • Fomenting creative communities and giving launchpads to new ventures is more valuable than pure content.

13. AI Engineering: Still the Path to Power

• Learning to Code Remains Crucial:
  • Levie and hosts agree—AI augments those who can code, not replaces the need entirely.
  • Software will be everywhere, so technical skills remain the supreme career lever.

Notable Quotes & Moments

| Timestamp | Speaker | Quote | |-----------|---------|-------| | 00:00 | Levie | "You don't write code. You talk to an agent and it goes and does it for you ... we are changing our work to make the agents effective." | | 04:15 | Host/Levie | "Every agent needs a box. If we can make the headline of this, I'm fine with this." | | 18:33 | Levie | "Most real life is extremely messy and like poorly named and duplic[ated], outdated shit." | | 22:08 | Levie | "How do I bridge the 50 million pages ... with the couple hundred ... I get to work with in that token window? ... that's why so much work is just search systems." | | 43:56 | Hoover | "Most companies are practically apprenticeships ... all that tacit knowledge is not written down, but it would have to be if you wanted to give it to an agent." | | 50:48 | Levie | "If we don't do this right, we're out of business ... you can just see how the AI tsunami could wipe you out ..." | | 59:05 | Levie | "To me, just building in public is just a natural thing. ... We deal with interesting problems. I tweet about them. ... My job is to try and connect all these things together and, and make it useful." | | 63:03 | Levie | "You get to be Spielberg ... is, you know, completely amazing ... but I still like the idea of like this is a camera and a person and a person that says action." | | 74:16 | Levie | "Things are only getting more technical, things are only going to get harder, and anybody in a technical position is in the best position to get agents deployed." |

Timestamps for Important Segments

• 00:00 – 05:00: Opening—Paradigm shift to agentic work, why "every agent needs a box"
• 05:00 – 12:00: Identity, agent permissions, and the unique problems of agent accountability
• 12:00 – 19:00: Why AI coding is further ahead than other knowledge work; challenges for the Fortune 500
• 19:00 – 25:49: Context engineering, context-rot, and model improvement anecdotes
• 29:44 – 35:39: Agent evaluation ("evals") at Box; measuring progress across industries
• 35:39 – 41:54: Team & product org structure, read vs. write agent workflows
• 45:06 – 50:23: Company knowledge graphs, wikis, and file system metaphors
• 50:44 – 55:44: Founder delegation, existential focus, keeping entropy at bay
• 59:05 – 65:36: Personal flywheel, content creation, reflections on AI in creativity/film
• 68:09 – End: Media as strategy, DevRel’s critical role, career advice for AI engineers

Key Takeaways

• Agent-driven workflows will soon be the default, but require massive changes to infrastructure and process.
• Data security, agent identity, and granular oversight are urgent, unsolved problems for enterprises.
• Context engineering—selecting, condensing, and curating relevant information for agents—is the new core AI challenge.
• Cultural, technical, and organizational adaptation lags far behind the advances in code-focused AI agents.
• AI will reward companies that are organized, well-documented, and ready to ‘feed’ agents; messy orgs will fall behind.
• Evaluation, benchmarking, and observability of agents are vital—every company will need a robust eval pipeline.
• Learning to code (and embracing technical skills) remains foundational for the age of AI agents and software abundance.
• Media, content, and DevRel have never been more strategic for product and talent reach in tech.

For the full set of show notes, charts, and reference documents: latent.space