OpenAI-Visa Agent Payments, Ona Acquisition, DeepMind Multi-Agent Safety — AI News Briefing

Top 7 Stories

1. OpenAI and Visa Partner to Let AI Agents Make Purchases

OpenAI and Visa have announced a landmark partnership that will enable AI agents to make online purchases and process payments on behalf of users. The integration will allow ChatGPT and other OpenAI-powered agents to complete transactions — from booking travel to ordering products — without requiring users to manually enter payment details each time.

The deal positions OpenAI at the center of what analysts are calling “agentic commerce,” where AI assistants don’t just recommend products but actually complete the purchase. VentureBeat reports that competing agent protocols from Google and others threaten to fragment the future of AI commerce, making this first-mover partnership strategically critical.

Visa’s existing merchant network gives OpenAI’s agents immediate access to millions of points of sale, while OpenAI provides Visa with a distribution channel into the rapidly growing AI agent ecosystem.

2. OpenAI Acquires Ona to Supercharge Codex Coding Assistant

OpenAI has agreed to acquire Ona, a startup specializing in persistent cloud execution environments, to bolster its AI coding assistant Codex. The acquisition gives Codex the ability to run code in persistent cloud sandboxes — a capability that’s essential for coding agents that need to maintain state across long-running development sessions.

According to CNBC and OpenTools, Ona’s technology will allow Codex to handle complex multi-file projects, run tests, and iterate on code without losing context between sessions. This addresses one of the key limitations of current coding agents, which often struggle with projects that span many files and require sustained context.

The move signals OpenAI’s deepening investment in AI-powered software development, a market that has become intensely competitive with Anthropic’s Claude Code, GitHub Copilot, and Google’s Gemini all vying for developer mindshare.

3. Google DeepMind Publishes Major Multi-Agent Safety Research

Google DeepMind has released significant new research on the safety challenges that emerge when millions of AI agents begin interacting with each other at scale. The research, covered by MIT Technology Review, explores scenarios where agent-to-agent interactions create emergent behaviors that no single agent’s designers anticipated.

The paper addresses what many researchers consider the next frontier in AI safety: not just making individual agents safe, but ensuring that ecosystems of interacting agents don’t produce harmful collective outcomes. With organizations rapidly deploying AI agents across their operations — a CTech report notes 52% of organizations using generative AI have already implemented agents — the urgency of this research is clear.

DeepMind’s work comes as the industry grapples with governance challenges for agent swarms, with Computer Weekly reporting that Google Cloud is separately unpacking the governance challenges of enterprise AI agent deployments.

4. NVIDIA Launches Cosmos 3 Foundation Model for Physical AI

NVIDIA has unveiled Cosmos 3, described as an “open frontier foundation model” designed specifically for Physical AI applications including robotics, autonomous vehicles, and industrial automation. The model is NVIDIA’s latest push to dominate the infrastructure layer for AI that interacts with the physical world.

Cosmos 3 is accompanied by a major collection of open-source agent tools and skills, as well as new Agent Skills for autonomous vehicles, robotics, and vision AI. NVIDIA is positioning the model as the foundation for the next era of physical AI research, with integrations across Amazon SageMaker AI and enterprise platforms.

The launch reinforces NVIDIA’s strategy of building a full-stack AI platform — from chips to models to developer tools — that locks in customers across the entire AI development lifecycle.

5. Google AI Mode Rolls Out Search Agents That Track Information

Google has begun rolling out AI Mode in Search, which introduces persistent Search agents that can track information over time and proactively surface updates to users. According to 9to5Google, the feature represents a fundamental shift from traditional search — where users ask questions and get one-time answers — to an ongoing relationship where AI agents monitor topics and notify users of relevant changes.

The feature also includes Google Cloud’s Open Knowledge Format, which converts scattered documents into Markdown files optimized for AI agent consumption. Together, these moves signal Google’s vision of search evolving from a query-response model to an always-on agent that works on the user’s behalf.

PCMag’s review of Gemini Spark — Google’s consumer AI agent — praised its capabilities but noted significant limitations, suggesting the technology still has room to mature before achieving mainstream adoption.

6. NVIDIA Hires Intel’s Former Top Lobbyist Amid AI Chip Wars

NVIDIA has recruited Intel’s former top government affairs executive as the AI chip wars intensify in Washington. The hire, reported by Memeburn and Electronics For You, comes as NVIDIA faces increasing scrutiny over its dominance in AI accelerators and as export controls on AI chips to China remain a hot-button policy issue.

The move signals that NVIDIA is investing heavily in its government relations capabilities at a time when AI policy decisions in Washington could significantly impact the company’s market position. TradingView reports that NVIDIA has also opened another AI chip door in China, suggesting the company is navigating a complex geopolitical landscape.

The hire also reflects the broader trend of tech companies investing in policy expertise as AI regulation accelerates globally, with the UK, EU, and US all developing sovereign AI strategies that intersect with chip supply chains.

7. Anthropic’s Claude Code Creator Manages Tens of Thousands of AI Agents

In a revealing interview with Fortune, the creator of Anthropic’s Claude Code disclosed that there are days when he manages tens of thousands of AI agents simultaneously. The statement provides a rare glimpse into the operational reality of running AI coding agents at scale and the infrastructure challenges involved.

Anthropic is also publishing research on coding agents in the social sciences, expanding the application of AI agents beyond software development into academic research. Meanwhile, Microsoft has released Fara-7B, an open-weight model that reportedly outperforms OpenAI on web automation benchmarks, adding another competitor to the agent landscape.

The sheer scale of agent management described by Anthropic’s engineer underscores a key industry challenge: building the orchestration, monitoring, and safety infrastructure needed to reliably operate thousands of autonomous agents concurrently.

Trend Watch

What to Watch

Agentic Commerce Standards: With OpenAI-Visa moving first and Google developing competing protocols, expect a battle over which agent payment standard becomes dominant. The winner will capture enormous transaction volume as AI agents increasingly handle purchasing decisions.

Multi-Agent Regulation: DeepMind’s research on agent ecosystem safety could become the foundation for new regulatory frameworks. Watch for policy proposals addressing agent-to-agent interactions, particularly in financial services where autonomous trading agents already pose systemic risks.

Coding Agent Consolidation: The Ona acquisition and Microsoft’s Fara-7B release suggest the coding agent market is entering a consolidation phase. Expect more acquisitions as companies race to build full-stack development environments with persistent execution, testing, and deployment capabilities.