Every AI vendor spent April promising you the same dream: autonomous agents — little digital employees that write code, close tickets, and answer customer emails while you focus on "strategy." Anthropic, OpenAI, Google — all three shipped agent platforms within two weeks of each other. Your LinkedIn feed looks like a hiring fair for robots.
Here's the thing nobody at the keynote mentions: your boss just Slacked you asking when YOUR team is deploying agents. The demos are gorgeous — an agent goes from bug report to pull request in minutes, the audience claps like they've never seen a bash script. But what actually happens after the demo ends and the cameras stop rolling?
Three launches landed back to back. Anthropic shipped Managed Agents on April 8 — cloud-hosted agent APIs (a way for your software to spin up and control agents remotely) at $0.08 per session hour. OpenAI updated its Agents SDK on April 15 with native sandbox execution — agents run code inside a sealed box so they can't break anything outside it. And Google Cloud Next kicked off on April 22 with a keynote titled "The Agentic Cloud," spotlighting ADK (Agent Development Kit), which shipped earlier this month. Google baked human-in-the-loop in as a first-class feature from day one — pause an agent mid-task, get a human to approve, resume.
Early adopters jumped in fast. Rakuten deployed specialist agents across five departments — product, sales, marketing, finance, HR — each going live in under a week. Asana's CTO reported shipping features "dramatically faster." Notion plugged Claude directly into workspaces for parallel task handling. And Sentry? Sentry went full send: their agent goes from flagged bug to opened pull request with zero human intervention. Fully autonomous. The vendor dream made real.
But here's the uncomfortable part. If you've followed independent research this month — and this channel has cited it enough that regular readers can recite the numbers — the pattern never changes. AI code ships 1.7× more defects. PRs rise 20% while incidents rise 23.5%. Developers delete a fifth of accepted AI code and heavily rewrite another 7%. Gartner predicts 40% of agentic projects die by 2027. More output, worse outcomes. Every single study. 😹
Andrej Karpathy called it on April 3 — before any of these three platforms shipped. "The industry is making too big of a jump and is trying to pretend like this is amazing, and it's not." Three weeks and three launches later, nothing proved him wrong.
This creates a structural gap between marketing and reality. Vendors compete on maximum autonomy because it demos beautifully on stage. But production data says the opposite: narrow scope beats broad capability. Read-heavy workflows (where agents analyze but rarely modify) beat write-heavy ones. Human checkpoints before any consequential action beat full autopilot. Even Sentry's "fully autonomous" success works precisely because bug-triage-to-PR is an inherently constrained domain — not because autonomy itself wins. 😾
Google might actually get this. Their ADK ships human-in-the-loop as the default path, not an afterthought. As SiliconANGLE's John Furrier wrote on April 20: "Features sit on top of platforms. Operating systems define the platform." The real competition isn't who builds the most autonomous agent — it's who builds the best control plane.
So when your boss asks about agents, don't forward the keynote clip. Ask one question about any platform: how easy is it to build a tightly constrained agent with explicit scope boundaries, read-only defaults, and mandatory human approval before anything consequential happens? If the answer is "well, you can configure that..." — walk away. If it's the default — you might have something. 😼
The smartest agent won't win the platform war. The most controllable one will. And that inverts every vendor's current roadmap priority. 🐈
