Agents: an operating‑model problem
Industry discussion has shifted from whether AI agents can be built to how they should be operated at scale, with hidden costs like multi‑model orchestration, integrations, governance and long‑term maintenance coming into view. Box CEO Aaron Levie’s notes from meetings with IT and AI leaders and a recent podcast both emphasise that agent projects are an operating‑model decision, not just a prototype exercise. (x.com)
The enterprise debate over artificial intelligence agents has shifted from “can we build one?” to “how do we run hundreds safely and cheaply?” (fastcompany.com) Box chief executive Aaron Levie said after meetings with more than 20 enterprise artificial intelligence and information technology leaders on March 19, 2026, that agents were “clearly the big thing” and that work was moving from pilots into production. He said coding remained the main use case in large companies, with other knowledge-work deployments starting to emerge. (fastcompany.com) In an April 2026 episode of the a16z Show, Levie said the hardest part was not the model itself but the operating work around it, including integrations and the fear from chief financial officers and chief information officers when agents start touching real systems. The episode’s published timestamps flag “CFO/CIO pushback” around integrations at 18:39 and a broader “AI adoption gap” at 42:53. (youtube.com) An agent is software that does more than answer a question in a chat box. It can search files, call tools, make decisions across several steps, and complete a task on its own, which is why companies now treat it more like a worker inside a system than a feature on a screen. (microsoft.com) That changes the job for information technology teams. Instead of picking one model and launching a demo, they have to connect company data, set permissions, monitor costs, decide when to use OpenAI or Anthropic or Google models, and keep the whole setup compliant over time. (boxinvestorrelations.com) Box’s own product launches show the stack getting heavier. In June 2025, Box added an Enhanced Extract Agent, a Model Context Protocol server, AI Studio updates, AI Units, and an AI Admin Console, and said those pieces were meant to help enterprises “operationalize AI” with more control. (blog.box.com) When Box launched Box Agent on April 2, 2026, it said the product used models from OpenAI, Anthropic, and Google and was built to work inside existing security, governance, and permissions controls. The company also added tools for administrators to build custom agents for multi-step business work. (boxinvestorrelations.com) Security groups are responding as if this is now a production problem, not a lab project. Microsoft on April 2, 2026 released an open-source Agent Governance Toolkit and said the December 2025 OWASP Top 10 for Agentic Applications had already defined risks such as tool misuse, memory poisoning, and rogue agents. (microsoft.com; owasp.org) Market forecasts point the same way. Gartner said on August 26, 2025 that 40 percent of enterprise applications would include task-specific artificial intelligence agents by the end of 2026, up from less than 5 percent in 2025. (gartner.com) Levie’s line is that companies are still early, but the workload has already changed. The prototype is now the easy part; the operating model is the project. (geekwire.com; fastcompany.com)
