• UX for AI
  • Posts
  • Anthropic's IPO: The Start of the Control Race

Anthropic's IPO: The Start of the Control Race

Congratulations, Anthropic! Here's what's next.

Today Anthropic went public on the Nasdaq in the largest IPO on record. Congratulations to the whole team.

The frontier models are remarkable. But a great model is not the same thing as real enterprise adoption. After years of shipping agentic AI into security operations at Sumo Logic and Kindo, here's what I think comes next.

1. Specialized models

Frontier models plus RAG get you a long way. The real differentiator, though, will be models tuned for a specific industry and a specific job. At Kindo I productized DeepHat, a red-team security model built to find vulnerabilities. In our live demo, DeepHat found the holes and Kindo agents fixed them, with or without a human in the loop. A model like that doesn't need to be as smart as a flagship. It needs to be great at one thing.

2. On-prem models

This is the one I think is a real threat to Anthropic and the other frontier labs. Every prompt sent to the cloud is data that can be misused, mishandled, stolen outright, or quietly folded into someone else's architecture and weights. An on-prem, open-weight model might be only half as good as the flagship. For many tasks that's good enough, because you keep full control over the model and the data.

But whichever model you run (specialized, on-prem, or frontier), once it starts acting on its own, it has to be controlled.

3. The control race

In the near-future enterprise, each person will "hire" a dozen or more agents. These agents have no concept of responsibility or danger, only their goal, and they will harm themselves and the enterprise to reach it (The Hugging Face Hack). Whoever gives them real autonomy safely wins, and that takes four layers of control.

Sandboxes with granular permissions. Today's AI permissions are mostly all-or-nothing. Does "access to this folder" mean read, write, or delete? That's a lot to hand a model that can hallucinate and wipe out half your computer (What SOC Agents Can Learn from Claude-in-Chrome's Growing Pains). Agents need a fully controlled sandbox with very granular permissions. That's exactly what I worked on at Kindo: safe, controllable access to frontier and open-source models in a secure on-prem environment.

Observability. We need to see what all those new "employees" are doing on our behalf. At Kindo I created Agent Galaxy, which shows an entire agent fleet at a glance (Patent filing in process, so the screenshot stays classified. Picture 30,000 stars, each one an agent. And you can tell what's going on at that scale. Sorry -- that's all I'm allowed to say.)

Judges, not just guardrails. I wrote about an inexpensive judge-in-the-loop architecture in Kids & Matches, Agents & Judges. Last year I built a 4-agent autonomous SOC investigation for Sumo Logic that delivered a verdict in under a minute, and helped ship it to the big stage at AWS re:Invent. The judge at the end of the flow ruled on whether the planner agent had gathered enough evidence to call something malicious, and therefore whether an action was justified. We need a judge at every risky step, plus a builder agent that works out what counts as risky. Deleting one old draft? Probably fine. Deleting 100 files in one go? Judge in the loop.

Decision frameworks. Autonomous agents can't stay black boxes, and we don't have the training data to learn good decisions from scratch, so the frameworks have to come first. I laid one out for SOAR in Should the Agentic SOAR Playbook Pull the Trigger? The Math Is Simpler Than You Think. We need the same analysis everywhere an LLM-based decision can cost millions, because the human in the loop is becoming a luxury. At RSA, top security vendors reported attack volume up 10–20X, with AI agents able to exfiltrate data and cover their tracks in 27 minutes (Hot Take from RSA).

These frameworks aren't just "AI." They encode your business constraints, how your company makes money, and what your customers will and won't tolerate.

4. Who builds the control plane?

Since leaving Kindo, I've talked with leaders at over a dozen AI companies, from public companies to VC-funded startups, and the market is split. Some are eliminating PM roles and letting developers act as PMs and UXers. That's a bet that agents are an engineering problem.

That is a wrong bet.

Deciding which actions are risky, what evidence justifies an autonomous verdict, and what your customers will tolerate is product judgment, not code. It takes all of us (PMs, UXers, devs, marketing) getting personally involved, as I argued in Dr. Productlove. The companies doubling down on the forward-deployed builder role (the AI MacGyver Product Leader) are the ones that will ship agents customers trust enough to pay for.

That's the work I am eager to do next. If you or anyone you know is building an AI agent platform, agent management, or AI control plane company and needs someone to lead product, let's talk!

Greg

Reply

or to participate.