Over the past two weeks there’s been another thread that’s less technical but that I think matters a lot: agents aren’t only maturing technically; the infrastructure for their commercial and economic side is starting to grow too.
First, a point of view. Sam Broner at a16z wrote an essay about agent payments. His core argument: agents won’t swipe credit cards; they’ll operate like businesses, with supplier terms and credit limits. He thinks programmable stablecoins may become the core payment rail for agent-native commerce. That sounds far off, but think about it: once an agent can run its own tasks, call APIs on its own and decide which service to use, it really does need a way to pay that doesn’t depend on a human authorizing each payment by hand.
Uniswap’s moves are on this line too. It opened a Developer Platform beta that lets developers generate API keys to add swaps and LP features. More interestingly, it launched a Uniswap Skill that lets agents integrate DeFi operations directly into their workflows. People have been talking up agents + DeFi for a while, but now someone has actually built the interfaces and tools.
Back to the more mainstream enterprise side: Google Cloud launched Data Agents in BigQuery and Vertex AI, letting teams deploy ready-made data agents or build their own with ADK. Enterprise analytics is moving from dashboards to agents that execute. OpenAI also expanded file support in the Responses API, so agents can now take in formats like docx, pptx, csv and xlsx. These sound boring, but they’re exactly what enterprise adoption really needs: can the agent read your company’s files, run your reports, connect to your data sources?
There’s an easily overlooked detail in the Claude Sonnet 4.6 update: it integrates Excel features for financial data sources like S&P, LSEG, Moody’s and FactSet. That’s not showing off model capability. It’s telling the financial industry: your agents can connect directly to your data now.
Security and trust are catching up too. Vercel ran automated security audits on skills, bringing in Snyk, GenDigital and Socket for independent reports covering more than 60,000 skills. Malicious skills get hidden, and risk levels are shown directly. It’s a lot like app-store review: once agent skills become an ecosystem, someone has to guard quality and security.
Gemini starting to run background tasks on Android is another signal. It isn’t about chatting with an AI; it lets agents carry out multi-step tasks like bookings and purchases directly from the OS layer. Google is embedding agent workflows right into the operating system, and once users get used to working this way, agents become part of the phone.
Anthropic’s acquisition of Vercept is also worth mentioning. Vercept works on perception and interaction, and it was brought in to push Claude’s computer-use abilities forward. Sonnet 4.6 already scores 72.5% on OSWorld, getting closer and closer to human level. Once agents can really operate software interfaces, complete transactions and fill in forms, and those abilities connect to payment systems, the agent economy is no longer just a concept.
My personal view is that starting in the second half of 2026, the way people talk about agents will shift from “what can it do” to “how much does it cost, who pays and how does it settle.” Once the technology is mature enough, business models and economic infrastructure are what really decide whether agents can scale.
I’d love to hear your take: what do you think the business model for agents will look like?