AI Agents Demand Enterprise Software Overhaul
Aaron Levie argues software must prioritize agent interfaces via APIs and CLIs, as coding agents excel at integrations humans struggle with, reshaping enterprise workflows despite CIO fears.
Agents Multiply, Forcing API-First Software Design
Aaron Levie emphasizes that with 'a hundred or a thousand times more agents than people,' enterprise software like Box must treat agents as primary users. This shifts focus from human UIs to agent-friendly interfaces: APIs, CLIs, or even ad-hoc code generation. Levie shares Box's rollout of an official Box CLI, enabling Claude (powered by Opus) to handle tasks like 'upload this entire folder from my desktop into Box' or 'process all these documents in this folder.' This 'blows your mind' in demos but reveals challenges at scale—e.g., 5,000 employees' agents hitting shared repositories could cause conflicts like concurrent file moves or deletes.
Steven Sinofsky counters that humans lack the 'algorithmic thinking' for complex flows, likening it to non-technical workers failing to flowchart marketing plans. Only one expert in 50 might fully document it. Martin Casado notes agents consolidate tools differently than humans, who limit to seven iPhone apps due to cognitive limits—agents face no such constraints.
"If you have a hundred or a thousand times more agents than people then your software has to be built for agents," Levie states, flipping the idea of marketing to agents as mere APIs. Instead, robust IDLs (interface definition languages) enable seamless interaction.
Coding Agents Succeed Where Knowledge Agents Falter
Levie highlights coding agents as a 'superpower' paradigm, seen in tools like Claw Cloud, OpenAI's super app, Perplexity Computer. These agents don't just read data—they code or API-call through workflows. Box's agent dynamically chooses: use existing skills/tools or write code on-the-fly for unique operations, covering 90% routine tasks plus edge cases.
This contrasts with struggling knowledge-work agents. Levie: agents navigate vast software surfaces (e.g., SAP help systems) better than humans, who bottleneck capabilities. Sinofsky agrees on consumption layers—AI fluidly handles PowerPoint bullets or Excel dual-axis graphs—but questions backend convergence to generic databases/APIs.
Casado describes early Nano Cloud bots needing broad integrations, but after days, they stabilize on essentials. Levie pushes back: agents enable 'integration on demand' at runtime, beyond pre-wired IT setups, gluing 75 systems in global supply chains—what CIOs have done manually for decades.
"The paradigm that appears to be taking off... is what if you give a coding agent access to your SaaS tools... it can actually code its way or uses APIs through whatever task," Levie explains.
Abstraction Layers Shift Workforce Skills Upward
Sinofsky draws historical parallels: his cousin, post-MBA, managed interns for spreadsheets she couldn't master—mirroring today's 'room of agents' coordinated by one systems thinker. Soon, users internalized spreadsheets, iterating M&A models from 2 to 30 times. Agents are at this 'Thanksgiving dinner' phase: rocket science now (e.g., Anthropic's growth marketer automating 5-10 silo jobs with Claude code), but skills will democratize.
Levie agrees jobs evolve: imagine infinite engineers beside each role. The viral Anthropic example required systems thinking, but agents will nudge non-experts. Sinofsky challenges with finite-demand jobs like $600 PC marketing vs. infinite-supply growth hacking.
"The job just moves up a rung... that's why I actually don't think anything about this is any different," Sinofsky asserts, predicting 'marketingish' agent skills collapsing layers.
"We're right at... when I'm using a spreadsheet already and... she's like 'I don't know why this is so hard'—two years later she's doing it," he adds on diffusion lag.
Enterprise Pushback: Integration Risks and Agent Permissions
CFOs/CIOs fear agents enabling rogue integrations, per Levie: six approached him post-talk, saying he's 'insane' for suggesting easier integrations. Humans+agents creating APIs between systems 27 and 38 could break systems of record. Read-only consumption scales now (large N, human oversight), but writes demand controls.
Solutions emerge organically: treat agents as employees. Users give personal agents API keys/email, now separate phone numbers/Gmail accounts with permissions (e.g., Gmail's RBAC). Levie: enterprises will provision agent identities, avoiding new control layers.
Casado notes loops in demos (e.g., endless nested directories), hitting limits. Sinofsky: backend systems like VA's 75 glued redundantly—perfect for agent integration.
"Unleashing not just the agents themselves but humans to do integration... 'please break my system of record'," captures CIO terror Levie encountered.
Diffusion Gap and Vastly Underestimated Economics
Sinofsky warns AI diffusion lags Silicon Valley hype: 'absurd to think you're going to vibe code your way to SAP'—domain knowledge isn't just data layers. Enterprises trail startups in adoption.
Levie: Wall Street underestimates by an order of magnitude. Compute budgets will dominate discussions; agent economics explode opportunities.
"The diffusion of AI capability is going to take longer than people in Silicon Valley realize," Sinofsky cautions.
"Everybody is trying to figure out the economics... off by at least an order of magnitude on how big the opportunity is," Levie counters optimistically.
Key Takeaways
- Design SaaS with agent scale in mind: prioritize APIs/CLIs over UIs, enabling 100x-1000x agent users.
- Leverage coding agents for integrations: they dynamically code/API-call, handling runtime queries beyond pre-built IT wiring.
- Expect abstraction shifts: like spreadsheets/interns, today's 'rocket science' agent orchestration becomes baseline skill in 2 years.
- Address enterprise fears proactively: provision agents as identities with permissions (e.g., dedicated Gmail/phone), starting read-only.
- Build hybrid agents: decide per-task between tools, APIs, or code—Box's model covers 90% routine + edge cases.
- Test at scale early: demos thrill, but simulate 10k hits/hour on shared repos to catch conflicts.
- Undervalue hype cautiously: domain depth slows diffusion, but economics are 10x larger than perceived.
- Use stories for intuition: Anthropic marketer (1→10 jobs), cousin's interns→spreadsheet mastery predict agent futures.