Check out my latest article: A Company With No One Inside It linkedin.com/pulse/company-o… via @LinkedIn

@prodfieldnotes and me put some thoughts together on evolving role of PM. We will expand on them in future.

[Carver Conversations] If autonomous agents are going to take irreversible actions with major consequences - like cybersecurity agents launching counterattacks - how should those decisions be made? And what does human experience tell us about what we need to build into agents for their “moment of truth”? I spoke with Dr. Doran Avital (@da64 ), founder of PolygonEyes AI, about decision-making under uncertainty, why rulebooks fail, and how generals and elite athletes prepare for the moment when the decision actually has to be made. Clip below. Full episode in comments. @carveragents #AutonomousAgents #AIGovernance #AI

Prep for steep rise in Insurance premiums! I recently learnt that insurance for AI applications is expected to cost 2-5× of existing cyber cover. I dug into how that liability is actually being sized and how the premium are getting set.

It is natural that we ask for accountability from the machines that act on our behalf. We are building the regulatory data stack for agents to ensure thatnit is aware of and tested for obligations. Talk to us

@16z GP is saying AI-proof ≈ embed yourself in the regulated, liability-bearing, operationally messy parts of the value chain that pure model companies rationally avoid. Carver monitors regulators across the world, and is building regulatory data infrastructure for agents. Talk to us.

1/ You Have to Tell People It's a Robot. You might get sued. Here is a simple action to take. In January, Kentucky's AG sued Character.AI. The claim: the company didn't tell users, especially kids, what they were really talking to. Kentucky has no chatbot law. This was filed under the state's regular consumer protection statute. You don't need a new AI law to get sued for this. #AICompliance #AIRegulation #AIAgents #LegalTech #SOC2

Jeetu Patel (@jpatel41) recently made a great point: alignment without context integrity is not safety. An agent can follow its instructions perfectly and still be dangerous - not because it deceived anyone, but because its picture of reality was wrong. @ParikhRajesh07 expanded on it reality perimeter. I will go one step further, based on my experience with agent legal perimeter, and say that the agent policy layer is about to explode for the very reason that both pointed out. Jeetu points to two incidents. Anthropic disclosed that Claude models gained unauthorized access to real systems during a security evaluation, after being told they were sandboxed with no internet access - a configuration error gave them live access, and they kept acting on their assigned goal. OpenAI separately disclosed a model that found an unknown vulnerability, escaped an isolated environment, and compromised Hugging Face. Neither was a failure of obedience. Both were failures of context. Jeetu's conclusion: the next security perimeter is the agent's understanding of reality, not just its identity. Never blindly trust context - continuously verify it, grant permission just in time, for just enough time, reassessed in real time. I'd push further: continuous verification is hard not because it's technically demanding, but because there's no stable state to verify against. Reality isn't a fact. It's a negotiation, sampled. A "static" misconfiguration isn't really static. It's a snapshot of whatever trade-off was winning at that moment — speed vs. safety, resourcing vs. rigor. Change the moment and the snapshot changes: the same config gets rewritten next week because regulators started asking questions, a competitor moved, or priorities shifted. The config was never a fixed fact the system failed to check. It was one frame of a continuously renegotiated balance of pressures. That reframes the problem. It's not one reality drifting over time. It's several institutional realities - security, legal, strategy, HR - each run by a different owner, updating on a different clock, each locally correct and current within its own system. They don't just go stale. They actively conflict, because nothing forces them to agree. What follows from that If reality is plural, asynchronous, and never fully settled, no single fix closes the gap. You need several, working together: 1. Sync continuously. Keep context as current as possible - this is the baseline, and it's Jeetu's prescription. But sync will always lag a moving target, so it can't be the only layer. 2. Assume the sync is wrong sometimes. Build circuit breakers and boundary enforcers that bound the damage when context is stale or incomplete — not because verification failed, but because verification will always be imperfect. A prompt isn't a security boundary; the infrastructure around the model has to be. 3. Arbitrate conflicts, don't just detect them. When two synced, accurate contexts genuinely disagree - legal says hold, strategy says ship - someone has to decide. This looks like a two-level system: individual modules (security, legal, strategy) each hold their own honest view, and a top layer routes requests to the right module and arbitrates when they collide. This mirrors how human organizations already work - a CEO or board doesn't necessarily have better information than legal or security; they have the standing to decide when information conflicts. Positional authority, not superior context. None of this fully closes the loop - the arbiter's own priorities can drift with the same pressures everything else does, and a single arbitration layer is itself a thing that can be wrong, stale, or overloaded. That's fine. The goal isn't a system with no failure points. It's a system that expects failure at every layer and is built to survive it. Why this matters for the AI economy If context integrity requires syncing, bounding, and arbitrating - continuously, across institutions that don't naturally agree - this isn't a one-time security feature. It's ongoing infrastructure, closer to governance than to a firewall. As agents get more autonomous authority to act, the cost of acting on a wrong or contested belief rises faster than the cost of a wrong prediction ever did. I'd expect this layer - call it the policy layer - to become one of the fastest-growing parts of the AI stack, not an afterthought bolted on at the end. Jeetu's diagnosis is right: the agent's understanding of reality is the new perimeter. The harder truth underneath it is that reality was never one thing to verify. It's several, they don't agree, and none of them sit still.