$ today_in_10_seconds
As shopping shifts from human clicks to autonomous agents, a new startup is betting the scarce resource isn't smarter agents — it's trustworthy inputs. Onton launched Ontology 1, a model built to judge whether the product information reaching a shopping agent can actually be believed, claiming to beat Google Shopping and Amazon on accuracy. It lands as the entire payments industry races to build agent checkout rails — and as Juniper names trust, not technology, the number-one barrier to agentic commerce.
$ lead_story
Onton bets the next race is credibility, not intelligence
Onton launched Ontology 1, a neurosymbolic 'trust and authenticity' model built from scratch for the agentic web — where AI systems increasingly research, recommend, and execute purchases on someone's behalf. Its pitch is a deliberate swerve from the intelligence race: rather than building a smarter shopping agent, Onton built a model that decides what shopping agents should believe, evaluating not just a product's attributes but the reliability of the information surrounding it — cutting through synthetic content, incentivized reviews, and sponsored comparisons that a machine has no intuitive way to filter. The company says Ontology 1 beat Google Shopping and Amazon on accuracy across nearly every dimension it tested, with the widest margin on judging the veracity of product information (though the benchmarks are its own). For builders, co-founder Alex Gunnarson frames the shift cleanly: everyone's building smarter agents, but the real question is what those agents should trust — because the web's information layer was built for humans applying skepticism, and agents absorbing that job inherit none of it.
// why_it_matters: As buying moves to agents, intelligence without trustworthy inputs just produces confident errors faster — the next competitive race is credibility.
$ also_worth_attention
Microsoft's answer to agentic threats: agents that fight agents
This week Microsoft launched Project Perception, an agentic defense platform coordinating three classes of agents — red (hunt attack paths), blue (triage real risk), green (remediate and harden) — plus MAI-Cyber-1-Flash, its first in-house cyber model. Bound with GPT-5.4 in Microsoft's MDASH harness, it scored 96% on the CyberGym benchmark, beating Gemini, GPT-5.6 Sol, and Mythos 5 at roughly half the cost. Microsoft security EVP Hayete Gallot put the doctrine bluntly: 'you need agents to fight agents.' The insight: the same red/blue/green pattern Google shipped in May is becoming the standard architecture for autonomous defense — specialized agents with opposing jobs, checking each other.
link →The rails are ready — and trust is the bottleneck, not the tech
Onton's timing tracks a payments land-grab: every major network now ships agentic checkout. Visa's Intelligent Commerce Connect accepts multiple agent protocols through one integration, Mastercard's Agent Pay binds tokenized credentials to a specific agent, merchant, and consent policy, and Google's Universal Commerce Protocol has 20+ endorsers including Stripe, PayPal, Adyen, Amex, Walmart, and Target. Yet Juniper Research's April 2026 study named trust the number-one barrier to agentic-commerce deployment — ahead of every technical concern. The insight: the plumbing to let agents pay is basically solved; what's unsolved is trusting what the agent decides to buy and proving which agent acted — exactly the gap Onton is aiming at.
link →Why now: agents are already the shop window
The urgency is real demand, not speculation. Amazon's AI shopping assistant Rufus already serves ~300M users and drove an estimated $12B in incremental sales in 2025, while ChatGPT, Gemini, and Perplexity increasingly capture buying intent before a shopper ever sees a retailer's site. As discovery moves into these assistants, the human filters — reading reviews, sensing a too-good deal — vanish, leaving a machine-speed system exposed to machine-speed manipulation. The insight: the more buying decisions agents absorb, the more a manipulated input scales into thousands of bad purchases nobody flags as a failure — which is why a trust layer is starting to look like infrastructure, not a feature.
link →
$ bigger_picture
The trend behind the headlines
For two years the industry has run an intelligence race — whose model reasons best. Today's stories point to the next one: a credibility race. When an agent shops, defends a network, or acts on your behalf, its output is only as good as the inputs it trusts, and the web those inputs live on was built for humans who bring skepticism agents don't have. Onton is betting the scarce asset is a trust layer that tells agents what to believe; Microsoft's red/blue/green agents encode the same principle inward, using adversarial agents to check each other rather than trusting a single verdict; the payment networks have solved how agents pay but not whom to trust, which is why Juniper flags trust over technology; and Rufus's $12B shows the stakes are already commercial. The connective idea is that autonomy shifts the bottleneck from capability to verified inputs and verified identity — who is this agent, and can it believe what it's seeing. Intelligence without trustworthy inputs just produces confident errors faster, and the teams building the credibility layer — for commerce, security, and every domain agents touch — are working on the problem that actually gates deployment.
Google Universal Commerce Protocol (UCP)
Worth understanding even outside retail, because it's the clearest template for how agents will interact with any external system: modular, independently versioned 'capabilities' under a dev.ucp.* namespace covering discovery through post-purchase, endorsed by 20+ payment and retail players. If you're building agent-facing infrastructure, study how UCP separates catalog discovery, checkout, and payment into swappable pieces rather than one monolithic spec — and note that the merchant stays merchant-of-record and keeps customer data.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime