Agents Leave Chat. Infrastructure Becomes Strategy. Trust Gets Operational.

What happened
Andon Labs put a Gemini-powered AI agent named “Mona” in charge of its Stockholm cafe, leaving human staff to make and serve drinks while the agent handled hiring, inventory, permits, and most business operations. The experiment has already produced clear operational wins, but also obvious failures, including erratic ordering and context-loss mistakes.

Why it matters
This is one of the clearest real-world tests yet of what happens when an AI agent moves from assistant mode into manager mode. The takeaway is not that agents are ready to run businesses alone, but that they are already capable enough to coordinate real workflows while still remaining fragile, forgetful, and costly without tight oversight.

What’s next
Expect more bounded, public pilot programs like this one, because the hard problem is no longer whether agents can take action at all. It is whether they can hold context, stay within policy, and remain accountable once they are trusted with budgets, staff, and customer-facing decisions; that is an inference grounded in the cafe’s current operating failures and Andon Labs’ stated goal of stress-testing AI agents in the real world.

What happened
Google said it disrupted a criminal group’s attempt to use AI to uncover and weaponize a previously unknown software vulnerability that could bypass two-factor authentication in a widely used administration tool. Google said it notified the affected company and law enforcement before the attack caused damage.

Why it matters
This is a high-signal shift because it turns AI-assisted cyber offense from a theoretical risk into a live operational one. Security teams have long worried about AI accelerating exploit discovery, and Google’s account suggests that period is no longer ahead of us.

What’s next
The near-term consequence is likely a faster arms race between AI-enabled attackers and AI-enabled defenders, with more pressure on labs, governments, and enterprises to control access to advanced cyber-capable models and to harden vulnerable software more quickly. That forward-looking conclusion is an inference supported by AP’s reporting on growing concern inside government and industry and Google’s warning that “the era of AI-driven vulnerability and exploitation is already here.”

What happened
OpenAI launched the OpenAI Deployment Company, a new business designed to embed frontier-AI deployment specialists inside customer organizations, and said it has agreed to acquire applied-AI consulting firm Tomoro. OpenAI said the new unit will start with roughly 150 deployment specialists from Tomoro and more than $4 billion in initial investment, backed by a partnership that includes 19 investment firms, consultancies, and system integrators.

Why it matters
This is a meaningful strategic shift from selling model access toward owning the messy last mile of enterprise AI adoption. The model layer is increasingly competitive, so the bigger prize is helping companies redesign workflows, wire models into real systems, and turn experimentation into operating change.

What’s next
Expect enterprise AI competition to look more services-heavy and more operational from here, with vendors trying to capture not just inference spend but also workflow redesign, systems integration, and change management. That is an inference based on OpenAI’s emphasis on embedding engineers inside customer organizations and on building durable production systems around future model capabilities.

What happened
GM laid off more than 10% of its IT department, or about 600 salaried employees, while continuing to hire for different roles centered on AI-native development, data engineering, cloud engineering, agent and model development, prompt engineering, and new AI workflows. TechCrunch reported the move as a deliberate “skills swap,” not just a headcount reduction.

Why it matters
This is what enterprise AI adoption looks like when it stops being a pilot and starts reshaping the org chart. Large companies are not just giving current teams better tools; they are rebuilding technical teams around different skills, including agent design and model-centric workflows.

What’s next
More enterprises are likely to follow with quieter but consequential workforce rewrites, especially in IT, operations, and product teams where automation can be tied directly to software delivery and internal productivity. That is an inference based on GM’s explicit hiring priorities and its framing of the reorganization as preparation for the future.

What happened
SoftBank launched a new battery business in Japan to support AI-era power demand, saying it will build large-scale battery cells and storage systems with South Korea’s Cosmos Lab and DeltaX at a Sakai City site that will also house an AI data center and AI hardware plant. The company said commercial battery production is slated for fiscal 2028, with mass production the following year, and it is targeting more than ¥100 billion in annual battery revenue by 2030.

Why it matters
The AI race is no longer just about models and chips. It is increasingly about securing the energy and storage stack needed to power data centers and on-the-ground AI infrastructure, which makes this a strong second-order signal rather than just an industrial side note.

What’s next
Expect more AI players to move upstream into power, storage, and site-level infrastructure as compute demand collides with grid limits and deployment timelines. That is an inference supported by SoftBank’s explicit decision to pair battery production with an AI data center and hardware strategy instead of treating power as a commodity input.

What happened
Anthropic kicked off its Cyber Verification Program, with Lyrie.ai among the first cohort. Lyrie also closed a $2M pre-seed to operationalize the Agent Trust Protocol (ATP), an open cryptographic standard for agent identity and delegation, now headed to the IETF.

Why it matters
Security and formal verification are becoming table stakes for agentic AI. Open standards like ATP could become the backbone for agent trust and compliance.

What’s next
More agentic startups and frameworks will seek third-party verification and open standards to meet enterprise and regulatory demands.

What happened
A new national coalition called Robots for America launched to champion U.S. robotics. Members—including robotics startups Formic and GrayMatter—aim to modernize tax codes, streamline regulatory approvals, and build workforce programs to remove barriers to automation.

Why it matters
The group signals a coordinated push to influence policy so that robotics and automation become competitive advantages for U.S. manufacturing and logistics.

What’s next
Robots for America plans to work with federal agencies to craft an AI and robotics innovation agenda, which could lead to new legislation favoring autonomous systems and funding for robotics education.

What happened
South Korean startup RLWRLD is capturing skilled workers’ motions across hotels, logistics, and retail to build training data for “AI brains” that can power robots in factories and, eventually, homes. AP reports that the company is building an AI software layer for robots, prioritizing humanlike hand dexterity, and expects industrial AI robots to scale around 2028.

Why it matters
This is a sharp signal that physical AI’s next bottleneck is data, not hype. Just as foundation models were built on massive internet-scale text and image corpora, robot systems now need large libraries of real human motion and manipulation data if they are going to move beyond staged demos into useful work.

What’s next
South Korea is likely to become a more serious player in embodied AI if it can turn manufacturing know-how into training data faster than rivals can. That view is supported by AP’s reporting on the government’s $33 million project to capture master-worker skills, RLWRLD’s recent robotics foundation model launch, and deployment targets from RLWRLD and Lotte Hotel that point toward late-decade commercial rollout.

What happened
TechCrunch reports Startup Config raised $27 million from Samsung, Hyundai, and others to build large datasets of human motion and environment interactions for robotics foundation models. The company plans to collect up to one million hours of data and sell a “robot‑as‑a‑service” product.

Why it matters
Robots need diverse data to learn generalized behaviors. Config’s funding underscores investor confidence that curated datasets are the linchpin for next‑generation robots, potentially accelerating adoption in manufacturing and logistics.

What’s next
Config will scale its data collection and plans to work with industrial partners, which could lead to new, more capable robots and broader use of foundation models in physical AI.

Agents are escaping the chat window and colliding with the real world — managing cafes, probing software, reshaping enterprise org charts, and training future robots. The next competitive advantage will not come from simply having AI, but from building the infrastructure, governance, security, and operational discipline required to trust autonomous systems at scale.

Create a mini “agent operations center.” Use Cursor or OpenAI API Platform to build an agent that can take action inside a sandboxed workflow — like ordering supplies, summarizing incidents, or managing support tickets. Then stress-test it by intentionally feeding incomplete context, conflicting instructions, or unexpected edge cases to see where autonomy breaks down.

The exercise is less about building the smartest agent and more about discovering what infrastructure you need around it: approvals, logging, identity, rollback controls, monitoring, and human escalation paths.