Agents Leave the Screen. Control Planes Rise. Trust Becomes Infrastructure.

What happened
OpenAI put Codex into the ChatGPT mobile app in preview, letting users monitor live threads, review outputs, approve commands, and steer work running on laptops, devboxes, or remote machines; OpenAI also says more than 4 million people now use Codex every week.

Why it matters
The real unlock is not “AI on a phone.” It is mobile supervision for long-running agents, which removes one of the biggest bottlenecks in agent workflows: waiting at your desk to unblock the next step.

What’s next
OpenAI says the feature is rolling out in preview on iOS and Android across all plans, while support for connecting phones to the Windows Codex app is still coming soon.

What happened
Fiserv launched agentOS for financial institutions, built with OpenAI on AWS Bedrock, and said it has already co-developed four agents with six bank partners for workflows such as commercial loan onboarding and report generation.

Why it matters
This is one of the clearest signs yet that agentic AI is moving into regulated core systems, not just sidecar copilots. Fiserv is packaging governance, auditability, kill switches, and human-in-the-loop controls as part of the product itself.

What’s next
Pilots are already underway, more deployments are expected this summer, and American Banker notes that other bank software vendors are likely to make similar moves.

What happened
OpenAI detailed new safety updates that let ChatGPT use context within and across conversations in rare, high-risk situations through narrowly scoped “safety summaries.” In internal testing, safe-response performance improved by 50% in single-conversation suicide and self-harm cases, and by 52% in harm-to-others cases across multiple conversations on GPT-5.5 Instant.

Why it matters
This is a meaningful capability shift from single-message moderation to pattern recognition over time. As generative AI becomes more persistent and personalized, safety increasingly depends on memory and context, not just filters.

What’s next
OpenAI says the current work is focused on self-harm and harm-to-others scenarios, but it may explore similar context-aware safeguards in other high-risk domains such as biology and cyber.

What happened
Anthropic and the Gates Foundation announced a $200 million, four-year partnership spanning grant funding, Claude usage credits, and technical support for programs in global health, life sciences, education, and economic mobility.

Why it matters
This is less about one philanthropic deal than about where enterprise-style AI deployment is heading next: sector-specific connectors, benchmarks, and evaluation frameworks embedded inside real institutions. It also expands the competitive AI battleground beyond commercial productivity into public-interest infrastructure.

What’s next
Anthropic says the partnership will produce connectors, benchmarks, evaluation frameworks, and other public goods, with the first education-related releases expected later this year.

What happened
Microsoft introduced MDASH, a multi-model agentic security system that coordinates over 100 AI agents to autonomously discover software vulnerabilities. MDASH has already found 16 real-world flaws in Windows networking and authentication, topping industry benchmarks and entering private preview for select enterprise customers.

Why it matters
This marks a leap in agentic AI for cybersecurity, shifting from research to production-grade defense. The breakthrough is not just in model performance, but in orchestrating swarms of agents to autonomously plan, test, and validate at scale—raising the bar for automated vulnerability discovery.

What’s next
Microsoft will expand MDASH access to more enterprise customers in June, while competitors like Anthropic and OpenAI race to deploy their own agentic security platforms. Expect rapid escalation in agent-driven cyber defense—and new questions about oversight and control.

What happened
eWeek reports that Figure said its Helix-powered humanoids ran an autonomous warehouse-style package loop for more than 17 hours and handled over 22,000 packages in a livestream, with robots swapping roles when batteries ran low.

Why it matters
Physical AI credibility will be won on endurance, throughput, and repeatability, not on stage tricks. This pushes the conversation closer to the metrics warehouses and factories actually care about.

What’s next
Even eWeek notes the run still needs outside validation, so the next meaningful milestone is independent proof that this performance holds up in day-to-day industrial deployments.

Agents are no longer waiting for humans to sit at a keyboard. From mobile supervision and banking control planes to autonomous cyber swarms and warehouse robots, the infrastructure for persistent, real-world AI execution is starting to solidify.

The next competitive advantage will come from operational trust: who can safely monitor, govern, coordinate, and validate autonomous systems at scale before they act in the real world.

Build your own “agent operations stack” in a weekend.

Use OpenAI Codex to run a long-running coding or research task from your laptop, then monitor and approve actions from your phone while away from your desk. Add a lightweight orchestration layer with Workato or LangChain, then simulate governance by adding approval checkpoints, audit logs, and kill switches inspired by Fiserv’s new agentOS model.

For the security layer, experiment with a multi-agent workflow using Microsoft Security Copilot or open-source agent frameworks to see how multiple specialized agents can coordinate vulnerability analysis, remediation, and monitoring—similar to Microsoft’s MDASH approach.