Back to Home

5 AI Agent Stories: Wallets, Routers, Court Wins

Another week, another pile of AI agent news. If you blinked, you missed a lot: the company that routes a fifth of the internet handed software agents actual wallets, NVIDIA shipped a router that picks models mid-task, a security startup banked $85 million on agent chaos, a federal appeals court told Amazon its shopping-agent ban was premature, and the industry started arguing about the enterprise brain. Here are the five stories worth your time.

1. Cloudflare hands AI agents a wallet

Cloudflare kicked off Agents Week by announcing Cloudflare Wallets and cloudflare.pay, giving agents deployed on its network a stable identity and a way to pay for things. Every account gets a unique web address that works as a persistent ID, and users can extend that identity to specific agents, so any business receiving a request can see exactly who authorized it.

The money side is a two-tier system. An Account Wallet receives, holds, and manages stablecoins, while Virtual Wallets are assigned to individual agents with guardrails built in from the start. Users set a spending cap, an approved merchant list, and a maximum transaction size the agent cannot exceed on its own. Handle registration opened this week, with full wallet access, including on- and off-ramping funds, rolling out in the coming months.

  • Stablecoin balances give agents a payment rail that works at machine speed.
  • Virtual Wallets keep spending limits at the per-agent level, not the account level.
  • Paired with the Monetization Gateway, Cloudflare says the two-sided agentic payment market is complete.

"The Internet is shifting from human-driven browsing to agent-driven commerce," CEO Matthew Prince said. "When an agent shows up at your door, you need to know who sent it. Cloudflare can give agents a face, a link to the human or organization that owns them."

2. NVIDIA's Switchyard routes tasks to the right model

NVIDIA open-sourced NeMo Switchyard, a model-routing library that lets agents pick the best model for each step of a task instead of sending everything to the biggest one. It ships with tuning-free routers, including an LLM classifier, a stage router that watches tool activity, and an escalation router that starts on a cheap model and upgrades when a session keeps failing.

The benchmark numbers are the story. LangChain tested the escalation router across 145 production-style agentic tasks and cut costs 74% versus a frontier-only baseline while sending just 7% of calls to the frontier model, at a roughly six-point accuracy tradeoff. Cognition wired the staged-routing method into Devin Desktop and delivered near-frontier results on its coding benchmark at about 28% lower mean cost.

The release landed alongside Nemotron 3.5 Lightning, a 30-billion-parameter open model built for long-running agent work that NVIDIA says runs on a single GPU and completes agentic tasks 30% faster than peers in its class.

3. Obsidian Security raises $85M to police non-human identities

Obsidian Security closed an $85 million Series D at a $1.1 billion valuation, led by Crescent Cove with participation from Greylock Partners and Menlo Ventures. The pitch: agents authenticate with API keys, OAuth tokens, and machine credentials, so the identity tooling built for humans simply does not apply.

CEO Hasan Imam says more than 65% of the company's enterprise customers have already granted agents access to data inside third-party SaaS applications, with 14 customers paying over $1 million a year. His warning to the industry is blunt: "The concept of detection and response doesn't work in the agentic world. Agents move in seconds, maybe milliseconds." Expect runtime enforcement rather than alerts to be the next security battleground.

4. Perplexity beats Amazon in court. The Ninth Circuit overturned a preliminary injunction that blocked Perplexity's Comet shopping agent from buying on Amazon. The court reasoned that the Computer Fraud and Abuse Act "contemplates access by a person," and that an agent executing a user's instructions is a tool, not an intruder. It also found the ban "would impair consumer choice and needlessly limit development of a nascent technology." Amazon says it disagrees and is evaluating its options, and the underlying lawsuit continues. Still, this is the first appellate signal that an agent acting for a user is not automatically a hacker.

5. The enterprise brain is the new agent. Agents keep stalling in production because they forget everything between sessions, so vendors are racing to build a shared memory and context layer underneath them. Microsoft shipped its IQ family, including Work IQ and Fabric IQ, and explicitly called it a context layer rather than a product. Glean is the pure play at roughly $300 million in annual recurring revenue, while UnifyApps co-founder Ragy Thomas, author of The Enterprise Brain, argues the model was always the easy part and architecture is the constraint. Snowflake measured a 20% accuracy gain and 39% fewer tool calls when agents got a governed context layer.

The through-line this week is that agents stopped being a model problem and became an infrastructure problem: who they are, what they can spend, which model they use, and what they remember. The winners will be whoever sells the plumbing.

Comments

No comments yet. Be the first to share your thoughts!