Over the past few weeks, three venture-backed startups announced funding rounds that, taken together, reveal something bigger than individual company success.
Orthogonal raised $4.3 million from Pantera Capital and Y Combinator to build discovery, orchestration and payments infrastructure for AI agents. The Prompting Company raised $6.5 million from Peak XV Partners, Base10 and Y Combinator to help products become discoverable by AI agents. Convey raised $38 million from Andreessen Horowitz to build AI teammates that execute enterprise workflows autonomously.
On the surface, these companies solve different problems. But together they point to a much larger shift.
AI agents are no longer being treated as software features. They're being treated as economic participants. And just as the internet required browsers, search engines, payment networks and cloud infrastructure before e-commerce became mainstream, AI agents are beginning to require their own infrastructure stack.
I believe we're watching that stack emerge in real time.
An AI agent cannot interact with a business it cannot find.
Historically, websites were designed for humans. Navigation menus. Landing pages. Contact forms. Search engines indexed pages for people. AI agents need something different. They need machine-readable descriptions of products, services and capabilities.
The Prompting Company is betting that discoverability will become one of the defining challenges of the agent economy. Their platform helps businesses create AI-optimised content so that when an agent asks "which CRM has the best API for small teams?" the answer includes their customers' products. They already host roughly half a million pages for clients including Rippling and Rho, and are collaborating with NVIDIA on next-generation AI search.
That feels like a reasonable bet. If agents can't find you, nothing else matters.
Finding a business is only the first step. An AI agent also needs to know what services are available. Can this provider enrich company data? Can that API verify an email address? Can another system process a payment?
Orthogonal is building infrastructure that allows agents to discover capabilities dynamically instead of relying only on tools they were originally given. Their platform currently supports more than 35 APIs, and through a single integration, agents can discover services, orchestrate workflows and complete payments.
Their CEO, Christian Pickett, put it bluntly: "Soon, there will be more agents than people online. Those agents will book flights, hire contractors, enrich customer data, conduct research, and complete transactions without a human in the loop."
Rather than hard-coding integrations, agents become capable of discovering new services at runtime. That is a significant architectural shift — and Pantera Capital's backing suggests it isn't just a startup hypothesis.
We're also seeing significant investment in systems that allow AI agents to execute work end-to-end.
Convey's $38 million Series A, led by Andreessen Horowitz, funds what they call "AI teammates" — agents that don't just assist with tasks but own outcomes autonomously. Their platform has already completed over one million hours of automated work for companies including NBCUniversal, Samsara and Unity. One large streaming service recovered 23,000 hours annually from reporting and ad operations workflows alone.
The distinction Convey draws is deliberate. An assistant helps you do work. A teammate owns the work. As their CEO Rohan Chopra explained, the companies winning right now are the ones removing the operational drag that prevents their teams from doing strategic work.
Increasingly, agents won't simply recommend actions. They'll perform them.
Suppose an AI agent discovers a law firm through a machine-readable directory. It understands the firm's services via an API. It gathers information from the client. Now it wants to make contact.
Who decides whether the enquiry should be accepted?
Who determines which information is mandatory before a conversation can begin?
Who explains why one enquiry qualifies while another does not?
Who records the reasoning behind that decision in a way that can be audited six months later?
Those questions are very different from discovery, orchestration or execution. They're questions of business policy, trust and governance.
As AI agents become a new class of customers — submitting enquiries, requesting services, initiating transactions — businesses will increasingly need infrastructure that allows them to receive AI-generated demand safely and consistently.
I believe this layer will become increasingly important over the next few years.
Businesses won't simply expose APIs. They'll expose machine-readable policies. Those policies may specify what information is required before an interaction can proceed, qualification criteria that must be met, routing rules that determine which team handles the enquiry, approval thresholds, escalation paths and audit requirements.
Instead of agents guessing how to interact with every business, businesses will publish their interaction policies in a format agents can discover and follow. That creates consistency for businesses while reducing ambiguity for agents.
This isn't entirely hypothetical. Financial services firms already document why they accepted or rejected customers. Legal and healthcare organisations already maintain intake governance. The difference is that these processes have historically been designed for human-to-human interactions. Extending them to agent-to-business interactions is a different engineering problem.
Another pattern is emerging alongside autonomous agents. As organisations rely more heavily on AI-assisted decisions, customers, regulators and internal governance teams increasingly expect those decisions to be explainable.
This trend isn't limited to any single jurisdiction. It's appearing wherever AI intersects with consequential decisions: customer intake, credit assessment, hiring, professional services engagement.
Questions like these will become increasingly common: Why was this enquiry prioritised? Why was another rejected? Which policy produced this outcome? Can we reproduce the same decision six months later?
The ability to answer those questions may become just as valuable as automation itself. Companies that can demonstrate transparent, auditable decision-making in their intake processes will have a meaningful advantage — both with customers and with regulators.
For decades, websites have relied on forms designed for humans. But AI agents don't interact like humans. They don't browse pages. They don't read marketing copy. They discover capabilities, exchange structured information and execute workflows.
The businesses that adapt first won't simply replace forms with chatbots. They'll expose structured, machine-readable ways for AI agents to interact with them — including the policies and qualification criteria that govern those interactions.
That transition feels comparable to the shift from desktop software to web APIs twenty years ago. The companies that published APIs early gained distribution advantages that lasted for years.
If recent funding announcements tell us anything, it's that investors increasingly believe AI agents require entirely new infrastructure.
Discovery. Capability routing. Execution. Payments.
I believe another layer is emerging alongside them. Business authorisation — the ability for organisations to define how autonomous agents interact with them, qualify work, apply business policy and explain every decision.
The companies building these layers aren't competing with one another. They're assembling different parts of the same future.
The interesting question isn't whether AI agents will become common. It's what infrastructure they'll need when they do.