Aaron Levie has been on a roll this week. Yesterday he was explaining why cloud bills are about to explode from multi-agent compute. Today he's making the case that the explosion won't even happen cleanly, because most enterprise workflows were never built for agents to plug into in the first place. The gap between "agents exist" and "agents work reliably in your company" is the story nobody in the sales deck wants to tell.
01
Deploying enterprise AI agents is harder than anyone is selling you
Box CEO Aaron Levie pushed back on the "just add agents" narrative with some pointed specifics: fragmented data, legacy systems that agents can't connect to, and institutional knowledge that lives in people's heads rather than documentation. His argument is that real enterprise AI deployment requires fundamental workflow redesign, not just a chatbot wrapper on top of existing processes. ---
Why it matters: Your vendor's AI demo ran on clean, structured data in a controlled environment. Your actual company has 14 years of spreadsheets, a CRM nobody fully trusts, and a process for approving invoices that only Karen in accounting fully understands. The gap between those two realities is where most enterprise AI projects are quietly dying right now.
Mistral AI announced Leanstral 1.5, described as a model focused on "proof abundance," though the blog content is thin on technical specifics. Given Mistral's track record with efficient, enterprise-ready open models, this is worth watching for benchmark details as they emerge. ---
Why it matters: Mistral has carved out a real niche as the option for teams that want capable models without full dependency on OpenAI or Google. If Leanstral 1.5 delivers on reasoning and stays lean enough for on-prem deployment, European enterprises with data residency requirements have a stronger option today than they did yesterday.
One company's AI tool is writing 65% of its product pull requests
Cat Wu shared that Claude Tag, an internal tool at her company, has reached a striking milestone: 65% of product pull requests are now landed by the AI. The rollout covers engineering, product, data, sales, and marketing, and she notes that security was built in from day one rather than bolted on afterward. ---
Why it matters: "65% of PRs" is the kind of number that will show up in board presentations for the next six months. The more important detail is the security-first framing. Teams racing to hit similar numbers without that design discipline are building a very different kind of risk into their codebase.
A small tip for anyone building with AI agents: group chats beat direct messages
Zara Zhang posted a short but practical observation for agent builders: routing agent communication through group conversations rather than one-on-one messages produces better results. The intuition tracks. Agents working in shared contexts can reference each other's reasoning, catch contradictions, and build on prior outputs rather than starting fresh every time. If you're architecting multi-agent systems and haven't tried this, it's a low-cost experiment. ---
The biggest applause line at an AI engineering conference was about mental health
Swyx, who organizes AI engineering events, noted that the moment that got the loudest response at this year's AI Engineers keynotes wasn't a product announcement. It was a conversation about men talking openly about their feelings and mental health while building companies in hypergrowth. Worth pausing on. The people building the infrastructure for the next decade of AI are human beings under significant pressure. The fact that this landed harder than any demo says something about the community right now.