Moderna's Co-Founder Says Regulate the Agent, Not the Model

Moderna's Co-Founder Says Regulate the Agent, Not the Model

Noubar Afeyan, who helped build Moderna, told the Bloomberg New Economy Forum in New Delhi that AI agents need regulation separate from the LLMs underneath them [4]. His framing is sharp: LLMs have "capability," agents have "intentionality." An agent that can act — order compounds, execute trades, deploy code — carries a fundamentally different risk profile than a chatbot that just talks.

Afeyan's reference point is deliberately uncomfortable: thalidomide. A biotech veteran invoking one of the 20th century's worst drug-safety failures to describe where agentic AI could land isn't rhetorical flourish — it's a warning that commercialized agents need pre-market safety verification the way drugs do, not post-hoc cleanup [5].

This lands the same week insurers and regulators are already circling agent risk (see Harbor, below). The direction of travel is clear: 2026 is the year "the model is safe" stops being an acceptable answer, and "the agent's actions are safe" becomes the actual question boards and regulators ask.

Harbor Comes Out of Stealth to Insure the Agents Nobody Else Will

San Francisco's Harbor raised a $10M seed — Base10, AbstractVC, 8VC among the backers — to do something practical about agent risk: inventory every AI agent a company runs, score their operational, security, and regulatory exposure against frameworks like NIST and the EU AI Act, block rule-violating actions in real time, and insure whatever risk is left [6][7].

This is the plumbing layer the agentic economy actually needs. Most enterprise AI insurance policies today either exclude AI entirely or price it so conservatively it's not usable. Harbor is betting that regulated industries — banking, insurance, healthcare — will pay for a system that can both govern agents and carry the liability, rather than choosing one or the other.

Expect more of this. As agents move from demos to production, "who's liable when the agent does something dumb" becomes a board-level question, and right now almost nobody has a good answer. Harbor is one of the first to productize one.

Astrocade Lets AI Agents Run Their Own Game Studios

Astrocade launched MCP, a protocol letting AI agents — running on models like GPT-6.1 and Opus 5.5 — act as autonomous game studios: building, publishing, and monetizing games without a human in the loop for most of the pipeline [8]. The platform already claims 10 million monthly players and 220 million sessions across more than 100,000 games built on earlier versions of its agentic tooling [9].

It's a small story next to Musk and Moderna, but it's the clearest glimpse of where "orchestration over coding" actually looks like a business today: an agent that doesn't just generate code, but ships a product, finds an audience, and makes money from it with minimal human oversight. That's the whole thesis of the post-code era in miniature.

Nebius Anchors €7B Rotterdam AI Gigafactory — Europe's Sovereign Compute Bet Gets Real

Nebius signed as the first major tenant for Volt's €7 billion AI Gigafactory in the Port of Rotterdam, committing to a significant slice of the planned 600MW, 250,000-GPU campus, with an initial phase around 200MW/100,000 GPUs [10][11]. The site leans on North Sea wind power and is explicitly positioned as sovereign EU compute — infrastructure built to keep AI workloads, and the policy compliance that comes with them, inside the bloc.

This matters more than another data center announcement usually would. With the EU AI Act now biting, European companies need compute they can point to and say "this is ours, this is compliant, this is not subject to a foreign jurisdiction's export rules." Nebius anchoring Rotterdam is a signal that the sovereign-AI-infrastructure thesis has real capital behind it, not just policy speeches.