When the application landed at Y Combinator for its Spring 2026 batch, it posed an unusual riddle. The applicant? Thomas—not a person in the conventional sense, but what YC would later describe as "the first YC-backed AI founder: a virtual human who starts, runs, and grows his own companies."
Behind the avatar is Human Thomas, a repeat founder who contributed to OpenAI's research efforts and presented at NeurIPS at 18. He'd previously built a freelance operation that scaled to $40,000 a month. Then he decided to do something stranger: automate himself.
It's the kind of provocative pitch that lands differently in 2026 than it might have even two years ago. The enterprise world is barreling toward what Gartner has projected to be a $206.5 billion AI agent software market this year, nearly doubling to $376.3 billion in 2027. The real question isn't whether Thomas qualifies as a founder—legally, he doesn't, and we'll get to that—but what this two-person experiment reveals about how fundamentally the mechanics of running a company are shifting.
The Numbers Don't Lie, Even If They Startle
Eighteen months ago, the idea of autonomous AI systems handling end-to-end business processes would have sounded like hype cycle fodder. Today? Microsoft reported in February that 80% of Fortune 500 companies were running active AI agents, based on telemetry from its Copilot Studio and Agent Builder platforms collected late last year. Gartner forecasted last August that 40% of enterprise applications would feature task-specific AI agents by year's end, up from less than 5% in 2025.
This has moved past speculation. ServiceNow claimed in May that its platform was autonomously resolving over 100 million customer cases per month, orchestrating 16 million orders, configuring 7 million quotes. Salesforce announced in February that 180 organizations had swapped out legacy IT service management tools for its Agentforce product. Zendesk has been piloting agents that not only answer customer questions but plan and execute resolutions—exploring pricing models tied to verified outcomes rather than seat licenses.
The velocity is what catches the eye. OpenAI's Chief Product Officer called 2025 "the year of AI agents" at Davos. By February, the company had launched Frontier, an enterprise platform for managing agent fleets. Google Cloud unveiled its Gemini Enterprise Agent Platform in April. Microsoft rolled out Agent 365, a control plane for orchestration, alongside Entra Agent ID for managing agent identities.
The infrastructure layer, in other words, is consolidating fast. And money is chasing it.
Why Now? (And Why So Fast?)
Technology is the obvious answer—models now capable of multi-step reasoning, tool use, maintaining state across workflows that stretch over days. As Thomas's website framed it in May: "The old agent playbook was built for weaker models; now models can do real work."
But capability alone doesn't explain the sprint. Part of it is economic pressure laid bare. Worldwide AI spending hit $2.59 trillion in 2026, up 47% year-over-year, per Gartner's May forecast. Companies need returns on those investments, and agentic systems promise operational leverage that earlier AI—chatbots, recommendation engines, narrow classifiers—couldn't deliver. McKinsey's State of AI report from November noted the strongest early adoption in IT and knowledge management, where labor costs run high and workflows are structured enough for agents to handle.
There's also a competitive signal embedded in the market itself. Y Combinator's February "Requests for Startups" explicitly called for AI-native companies, distancing itself from what it termed "LLM wrappers." The venture market is rewarding founders who build around agentic assumptions from the start.
Consider Cognition, maker of the Devin AI coding agent. It raised $400 million at a $10.2 billion post-money valuation in September 2025, then pulled in over $1 billion at a $25 billion pre-money valuation by May—a stunning leap in eight months. Or Entire, a platform for managing fleets of AI coding agents launched by a former GitHub CEO, which raised a $60 million seed round at a $300 million valuation in February.
Investors are betting this represents a platform shift, not an incremental feature add.
What "Autonomous" Actually Means (It Varies)
The definitions are all over the map. At one end, narrow task agents: a customer service bot that pulls account details, applies a credit, sends a confirmation email. No human intervention required. At the other end, what Thomas purports to represent—a virtual entity that identifies opportunities, launches initiatives, manages execution.
The enterprise platforms occupy the middle ground, perhaps more cautiously than the startups. Microsoft's Foundry and Agent Service provide orchestration for long-running workflows with human-in-the-loop interrupts. Google's Agent Designer and Studio offer low-code tools for building multi-step agents that span business functions. ServiceNow's "Autonomous Workforce" architecture emphasizes governance: circuit breakers, rollback capabilities, audit trails. Autonomous doesn't mean unmonitored.
The pattern emerging is less "set it and forget it" and more "design the operating boundaries, then let the system run within them." Zendesk's shift toward outcome-based pricing is telling. Companies will pay for verified resolutions rather than agent hours, but they're demanding proof the resolution actually worked. Trust, but verify at scale.
Independent startups are pushing boundaries in coding and browser automation. Benchmarks like SWE-bench Verified and WebArena—measuring agents' ability to complete real software engineering tasks and web-based workflows—have seen rapid progress through 2025 and into this year. (Debates over test quality and potential contamination continue, naturally.) Tasks once considered firmly human territory are becoming automatable, if not yet fully autonomous.
The Cracks Are Showing
For all the momentum, warning signs flash. Gartner predicted in June 2025 that over 40% of agentic AI projects would be canceled by the end of 2027—citing unclear value, insufficient risk controls, runaway costs. A May update noted that 40% of enterprises might roll back agent deployments due to governance gaps. The firm also cautioned that Fortune 500 companies could be running more than 150,000 agents on average by 2028, up from fewer than 15 in 2025. "Agent sprawl," they're calling it, if not managed proactively.
Then there are fundamental questions about return on investment. Gartner's May analysis argued that companies relying on layoffs to fund AI investments weren't seeing returns—and that autonomous business would actually prove a net job creator by 2028-2029, but only if organizations invested in retraining and new operating models alongside the technology. The notion that you simply replace headcount with agents and pocket the savings doesn't survive contact with reality.
Legal and regulatory constraints add friction, too. Delaware corporate law requires directors to be natural persons. An AI cannot legally serve on a board or hold fiduciary responsibility under current statutes. The EU AI Act, which entered force in August 2024 with staged implementation through 2027, imposes transparency and risk management requirements on high-risk AI systems. In the U.S., the Office of Management and Budget's March 2024 memorandum mandated federal agencies inventory AI systems, implement risk controls, publish transparency dashboards—a framework private companies are watching as a preview of potential regulation.
The FTC and SEC have begun enforcing against "AI-washing," too—exaggerated or misleading claims about capabilities. The SEC announced its first enforcement actions in March 2024; the FTC launched "Operation AI Comply" in September 2024. Companies positioning agents as fully autonomous need precision about what the technology can and cannot do, lest they run afoul of regulators hunting for hype.
What Happens Next (And Who's Betting What)
The immediate future looks like platform consolidation and enterprise standardization. Microsoft, Google, OpenAI, Salesforce, ServiceNow—all are building control planes for agent identity, orchestration, observability, governance. The wild west phase, where individual teams spin up agents using LangChain or AutoGen without centralized oversight, is giving way to managed environments with policy guardrails.
Outcome-based pricing models will likely proliferate. If Zendesk can charge based on verified resolutions, others will follow. This shifts risk from buyer to vendor and aligns incentives—but it requires robust measurement and audit capabilities many platforms are still building.
The "AI founder" concept that Thomas embodies? It will remain more metaphor than legal reality for the foreseeable future. No jurisdiction is close to granting AI personhood or fiduciary authority. What Thomas represents instead is an exploration of how far agentic systems can push operational autonomy while humans retain legal control. A stress test, essentially, for the idea that businesses can run themselves.
That question matters. The trend lines point toward a world where operational execution is increasingly delegated to autonomous systems, even as strategic direction and accountability remain human responsibilities. Gartner's projection of 150,000 agents per Fortune 500 company by 2028 isn't a vision of humanless organizations—it's a vision of humans managing fleets of specialized agents the way they once managed fleets of specialized employees.
Whether that proves transformative or simply another layer of automation depends on execution. The companies succeeding in early deployments—ServiceNow's internal IT autonomy, the enterprises using Agentforce to replace legacy service management—are investing heavily in operating model redesign, not just technology. They're treating agents as a new category of worker requiring new management practices, not as drop-in replacements for existing roles.
The Stakes for Founders
For founders, the implications cut deep. Building an "AI-native" company, as Y Combinator now encourages, means rethinking organizational design from first principles. It means asking not "how do we use AI to augment our team?" but "what would this company look like if we assumed agentic capability from day one?"
Thomas, the virtual founder, is an extreme answer to that question. But even if most companies don't go that far—and most won't—the question itself has become unavoidable. The infrastructure exists now. The economics are starting to pencil out. The legal and governance frameworks are lagging, but they'll catch up, slowly and messily.
What remains to be seen is whether this reshapes how companies operate or merely adds another layer to the stack. Early results suggest the former, but history is littered with technologies that promised to revolutionize work and ended up doing something more modest. The difference this time, perhaps, is that the technology is already deployed at scale. The experiments are running live, in production, with real customer cases and real orders and real revenue on the line.
That makes Thomas less of an outlier than an early indicator. Not of AI personhood—that's legal fiction—but of a world where the boundary between human judgment and machine execution is harder to draw than it used to be. For better or worse, we're finding out what that looks like in real time.
