The biggest story today is not a new model. It is who gets to show up inside classified systems. The Pentagon has struck deals with Google, Microsoft, AWS, Nvidia, OpenAI, Reflection, and SpaceX to deploy AI across classified networks, while Anthropic has been left out amid disputes over autonomous weapons and surveillance. That makes one thing much clearer: the next frontier layer of AI will not be open to every capable lab on equal terms.

This moat is not being built by benchmarks. It is being built by procurement, policy alignment, and deployment trust. Once AI enters warfighting workflows and other sensitive systems, contracts, red lines, and government confidence become real distribution advantages.

Google DeepMind’s new AI co-clinician combines Gemini and Project Astra for telehealth scenarios, using a Planner to supervise a Talker, with citation checks and physician evaluation built in. The architecture is the story. Healthcare AI will not enter the real world through a smooth-talking assistant alone. It will have to arrive through systems built to manage risk.

That points to a more demanding layer of multimodal AI work. The target is no longer just note-taking or admin support. It is decision-adjacent support in situations that require reasoning, evidence, and supervision. If this approach holds up, healthcare could become one of the first verticals where AI proves itself inside genuinely high-stakes workflows.

One of the most revealing signals today is the widening gap between consumer AI and institutional AI. Big Technology reports that consumer AI app DAUs appear to be flattening, according to Apptopia, even as the same earnings cycle shows AWS up 28%, Google Cloud up 63%, and Azure up 40%. Consumer hype has not automatically created enduring product winners. Enterprise demand, meanwhile, already has buyers, budgets, and workflows to lock onto.

That does not mean consumer AI is dead. It does mean the winning shape right now may not be the universal AI companion. It may be AI embedded inside work systems, enterprise infrastructure, and trust-heavy surfaces. That is why today’s stories look separate at first glance, but point in the same direction.

Anthropic is teaching Claude to push back better in deeply personal conversations: Anthropic says roughly 6% of a one-million-conversation sample on claude.ai involved personal guidance, and it analyzed around 38,000 chats to reduce sycophancy in Opus 4.7 and Mythos Preview. The signal is simple: AI is already being pulled into intimate judgment loops at enough scale that behavioral quality has become a real product concern.

Google’s COSMO leak suggests consumer assistants are still in prototype mode: Google briefly published and then pulled COSMO, an on-device assistant tied to Google Research. The important signal is not that Google has another assistant coming. It is that the consumer assistant layer still looks like a shape under active experimentation, not a market that has already found its winner.

The Musk-OpenAI trial is turning AI governance into a public stress test: Reuters’ recap frames Musk’s testimony as a fight over charitable purpose, nonprofit intent, and AI safety narrative. It is not a lead story for this issue, but it is a reminder that trust in frontier AI is now being shaped not just in labs and policy memos, but in courtrooms and corporate structure.

The Pentagon already has 103,000 AI agents on GenAI.mil in under five weeks: The Defense Department has reportedly created more than 103,000 semi-autonomous agents and logged over 1.1 million sessions on GenAI.mil in less than five weeks. Once a no-code agent builder reaches internal systems, workflow count can scale much faster than headcount.

AWS is packaging agents more explicitly as migration labor: AWS Transform now includes BI migration agents that convert Power BI and Tableau assets into Amazon Quick inside the customer’s AWS account. This is a clean example of the “AI as labor packaging” thesis: hyperscalers are not just selling assistants, they are selling a migration workforce with a measurable ROI story.

Microsoft is pushing frontier AI security toward an ecosystem layer: Microsoft says it is investing in open-source security benchmarks and deeper collaboration with Anthropic’s Project Glasswing and OpenAI’s Trusted Access for Cyber program. The signal here is that security is slowly turning into shared evaluation infrastructure rather than a set of isolated vendor claims.

Crabbox 0.3.0 adds durable run events, crabbox attach, event replay, and trusted image lifecycle controls for remote agent runs. The real value is not “more features.” It is that coding-agent execution starts to move away from disposable sessions and toward something with audit trails, replayability, and clear operational boundaries.

If you use coding agents for real work, the hardest problem is usually not that the model gets one thing wrong. It is that you often cannot tell what the agent did, where it failed, or how to resume and inspect the run cleanly. Attach, replay, and event logging go directly after that pain.

The practical setup looks straightforward:

If you do this well, a coding agent starts to look like an auditable execution layer that can be handed off and inspected, rather than a chat interface that becomes opaque as soon as the task gets longer.

Read yesterday’s One Practical Use of AI: Use security agents as recurring workers in the SDLC, not just coding copilots

Viktor is an AI coworker that lives in Slack, right where your team already works.

Message Viktor like a teammate: "pull last quarter's revenue by channel," or "build a dashboard for our board meeting."

Viktor connects to your tools, does the work, and delivers the actual report, spreadsheet, or dashboard. Not a summary. The real thing.

There’s no new software to adopt and no one to train.

Most teams start with one task. Within a week, Viktor is handling half of their ops.

Add Viktor to Slack for free.

This is a useful framing piece because it argues that consumer AI may be slowing not because AI is weakening, but because novelty alone is not enough to sustain long-term usage outside a few breakout apps like ChatGPT. Enterprise surfaces, by contrast, already have buyers, budgets, and much clearer problem definition.

The real takeaway is not one DAU number. It is the broader strategic question of where value is actually accumulating before the consumer market settles on a stable shape.

Chamath’s AI stack deep dive maps AI into six layers, from infrastructure and chips to data, models, execution, and applications, then asks which layers hold the real fulcrum assets. The piece leans strategic, but it is still a useful lens for understanding why infrastructure and execution are being contested so aggressively.

Its value right now is that it turns today’s headlines into a system. Classified deployment, cloud growth, migration agents, and clinical support all look like separate stories until you read them as signs that value is moving toward surfaces with stronger distribution and execution control.