Trust Is the New Infrastructure: How the Agent Economy Is Building Its Safety Layer

Something is clarifying in the agent economy this week, and it isn't a single product launch or a funding round — it's a meta-pattern. Across infrastructure, payments, security, and identity, the dominant theme is trust: who can verify that an agent is acting correctly, on whose authority, and within what limits. The pieces are arriving from different directions, but they're assembling into the same picture.

The most striking signal came from The Decoder this morning: Anthropic's Opus 5, running in what the company calls Auto Mode, achieved a zero percent prompt injection success rate across 129 test scenarios. That's a bigger deal than it might sound. Prompt injection — where malicious content in a webpage or document hijacks an agent's instructions — has been the single most stubborn security problem holding back browser-based agents in production. It's why enterprises have been wary of giving agents access to real systems with real consequences. If Opus 5's resistance holds up under more adversarial testing, it potentially unlocks the browser agent category in a way that no governance policy or sandbox could.

But model-level hardening is only one layer of the trust problem. The deployment layer remains a mess. OpenAI this week introduced Presence, a managed service that delivers enterprise AI agents bundled with OpenAI engineers and integrators — a tacit admission that self-serve agentic AI is too ungoverned for most enterprise buyers. The governance gap is real: independent surveys suggest 41–44% of organizations have deployed AI agents without basic human-in-the-loop controls, and more than half lack kill switches or purpose-binding constraints. Presence is OpenAI charging a premium to solve its own customers' inability to govern what OpenAI sold them, but the market clearly needs it.

The commerce layer is moving toward standardization just as quickly. Adyen this week shipped support for three distinct agent payment protocols simultaneously — the Universal Commerce Protocol, Agent Payments Protocol, and OpenAI's Agentic Commerce Protocol — with Visa, Mastercard, American Express, and Salesforce listed as early participants. The payment rail war that Natural's $30M raise illustrated last week is now entering a phase where the largest acquirers are betting on interoperability rather than winners. That's a sign the market is maturing: when the infrastructure players start building adapters instead of backing horses, it means they expect several protocols to coexist, not one to win.

Then there's the identity layer, which might be the strangest and most telling thread. The World Foundation just raised $52.5M to scale Sam Altman's biometric proof-of-human network — a system designed specifically to let agents and humans coexist in digital environments where the distinction matters for authorization. Franklin Templeton's research team called agentic AI "crypto's killer use case" this week, for the same underlying reason: autonomous agents moving money need cryptographic identity, not password-protected logins. The infrastructure for verifying that an agent is authorized, bounded, and traceable is attracting capital from every direction at once.

The through-line across all of it is simple: the agent economy has been building capability for two years. This week's news suggests it's now building the layer that makes capability safe enough to actually deploy at scale. Trust isn't a feature you add at the end. It's turning out to be the infrastructure.