When Y Combinator announced its Spring 2025 batch last June, one number jumped out: reports varied between 67 and 70 of the 144 accepted startups—nearly half—were building autonomous AI agents. Not chatbots. Not recommendation engines. Full-fledged systems designed to plan, execute multi-step tasks, and learn along the way, all without someone constantly looking over their digital shoulder.
It's the kind of concentration that makes previous startup pile-ons look restrained. Remember when every pitch deck claimed to be "Uber for X"? This feels different, though—less about founders chasing a hot category and more about capital repositioning for what several major analyst firms are calling an inevitable shift in how enterprises operate.
The question hanging over Silicon Valley isn't whether agentic AI will happen. Platform giants are already embedding agent builders into their core infrastructure. Gartner predicted last August that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from under 5% in 2025. A separate study from Cisco and Omdia, released in February, found that 80% of executives believe agentic AI will prove essential to survival by 2027. Investors appear to be reading the same forecasts—and betting that nimble startups can outmaneuver slower-moving enterprise vendors.
The real question is whether nearly half a YC batch betting on agents represents foresight or hubris.
Beyond the Chatbot
Agentic AI marks a departure from systems that merely respond to prompts. These are tools meant to handle complex, multi-step workflows autonomously: booking travel itineraries, debugging code across repositories, triaging customer complaints and escalating only the thorny ones. Gartner frames them as "task-specific AI agents" that will eventually evolve into cross-application, multi-agent ecosystems—a vision that sounds ambitious until you realize how quickly the infrastructure to support it has materialized.
Andrew Ng, the AI researcher and investor, has been advocating for "agentic workflows" since 2024. His argument, which has gained traction in technical circles, holds that design patterns like reflection, planning, and tool use can deliver larger practical gains than simply throwing more powerful models at problems. It's an architectural thesis, not just a capabilities one, and it seems to be landing with founders and venture capitalists alike.
CB Insights tracked the field expanding from roughly 300 to over 400 "promising" agentic AI startups across 16 categories between October 2025 and early this year. That's a significant clip, especially given how nascent the technology remains. Meanwhile, the hyperscalers have moved with unusual speed. Microsoft rolled out Copilot Studio agents at Ignite 2024, then layered in Entra Agent ID and broader governance tools at Ignite 2025. Google launched its Vertex AI Agent Builder in April 2024 and began charging for the Agent Engine last March. AWS brought multi-agent collaboration in Amazon Bedrock to general availability around the same time.
Salesforce, never one to miss a platform shift, announced Agentforce in September 2024 and scaled it up to Agentforce 360 at Dreamforce in October 2025. CEO Marc Benioff positioned Slack as an "agentic OS"—a phrase that drew skepticism from some observers but reflected the company's urgency. By March, Salesforce had pushed Agentforce Contact Center into general availability, targeting the customer service trenches where agents might actually prove their worth.
The Technical Substrate
The capabilities undergirding this wave arrived in rapid succession, sometimes faster than the industry could absorb. Anthropic introduced a "computer use" tool in October 2024, letting Claude automate graphical interfaces—the kind of thing that sounds minor until you realize it means an agent can now navigate software built for humans. OpenAI followed in March 2025 with a Responses API that exposed similar functionality alongside developer tools for tracing and evaluation. GitHub, meanwhile, rolled out Copilot "agent mode" between February and May 2025, enabling long-running edits across multiple files. By February of this year, GitHub was supporting additional coding agents beyond its own, including Claude.
Open protocols emerged to knit these systems together, though whether they'll hold remains an open question. Google announced the Agent2Agent (A2A) protocol last April, with Microsoft committing to support it in Azure AI by May. Anthropic's Model Context Protocol (MCP), introduced in November 2024, has gained adoption across several platforms. LangGraph, an open-source framework for building durable agents, reached general availability in October 2025. Microsoft re-architected AutoGen—its multi-agent framework—with version 0.4 in February 2025, then began unifying it with Semantic Kernel into something it's calling the Agent Framework.
But—and there's always a but—the gap between technical possibility and operational reality remains uncomfortably wide. Camunda's "State of Agentic Orchestration & Automation 2026" report, published in January after surveying 1,150 leaders, found that while 71% claimed they use AI agents, only 11% of agentic use cases actually reached production in the prior year. Seventy-three percent reported a gap between their agentic AI vision and what they'd managed to deploy. The top obstacles? Business risk, cited by 84% of respondents. Transparency and compliance concerns weren't far behind at 80% and 66%, respectively.
Gartner's own survey from September 2025 echoed the caution: only 15% of IT application leaders were considering, piloting, or deploying fully autonomous agents as of mid-2025. Seventy-five percent had some form of AI agents in play, sure—but not necessarily the autonomous variety that investors are getting excited about.
The Early Movers
Some companies, though, have pushed past pilots and into the deep end. Cognition Labs, the startup behind Devin—an autonomous software engineering agent—raised $400 million in September 2025 at a valuation around $10.2 billion. The round, led by Founders Fund, came alongside reports of Devin pilots at Goldman Sachs. It's a staggering number for a relatively young company, reflecting investor confidence that developer productivity is a category where agents can gain traction quickly. Tasks are well-defined, outcomes measurable, and the market—enterprise software teams—has shown willingness to pay for tools that genuinely boost output.
Hippocratic AI, focused on healthcare, raised $126 million in a Series C last November at a $3.5 billion valuation. The company claims more than 115 million clinical interactions across over 50 health systems, payers, and pharmaceutical companies in six countries. It's an eye-popping scale claim for a startup, though verification remains difficult from the outside. The emphasis on safety-focused healthcare agents suggests vertical specialization might be one way through the governance and compliance thickets that trip up more horizontal approaches.
Customer service has become another early testing ground, perhaps inevitably. Zendesk launched an autonomous support agent in October 2025, claiming it could resolve 80% of support issues—though that's vendor guidance, not third-party validation, and should be taken with appropriate skepticism. Intercom rolled out Fin Voice in March 2025, extending its Fin AI Agent into voice-based interactions. Salesforce's Agentforce Contact Center, now in general availability, positions agents as a way to handle routine inquiries while routing complex cases to humans. Whether customers actually prefer interacting with agents over humans remains an unanswered question, one that companies seem willing to test at scale.
Voice automation is drawing capital as well. Synthflow AI raised a $20 million Series A last July to build enterprise voice agents. Retell AI, which raised a $4.6 million seed in May 2024, received an OpenAI partner spotlight in June 2025 for GPT-4o-powered voice agents. Both reflect investor interest in applying agentic workflows to business process outsourcing and contact center operations—industries where labor costs are high and tolerance for experimentation is growing.
Even management consulting firms are experimenting internally. Reports from January indicated that McKinsey operates with roughly 25,000 "AI agents" alongside approximately 40,000 human employees, with leadership expecting broader teammate-agent pairing by year's end. The figure, which emerged through media coverage, should be treated as a reported claim rather than verified fact. But it signals how early adopters—organizations with resources to burn and appetite for risk—are testing these systems at scale, learning what breaks and what holds.
The Regulatory Reckoning
The industry faces a regulatory inflection point this year that could reshape the competitive landscape. The EU AI Act's high-risk AI obligations take effect on August 2, requiring technical documentation, risk management, logging, and human oversight for many autonomous systems. General-purpose AI transparency obligations already kicked in last August. U.S. federal buyers and vendors serving the public sector must align with OMB guidance issued in 2024, which mandates impact assessments, inventories, and strengthened procurement clauses for AI features.
These requirements will favor platforms with built-in identity, observability, and policy controls—capabilities Microsoft, Google, and AWS are embedding now, likely with an eye toward the compliance premium they'll be able to charge. Startups without deep pockets for governance infrastructure may find themselves squeezed, either forced to partner with hyperscalers or to focus on niches where regulatory burdens are lighter.
Security remains an open question, one that keeps CISOs up at night. The OWASP Top 10 for LLM applications highlights risks like prompt injection, data exfiltration, and tool misuse—all of which become more acute when agents can execute actions across systems without human checkpoints. NVIDIA introduced NIM microservices for NeMo Guardrails in January 2025 to address multi-agent safety controls. Microsoft added Entra Agent ID last December to provide identity management for agents. Academic researchers published analyses in late 2025 and early this year on agent penetration testing and MCP security, proposing protocol hardening. Whether these measures will prove sufficient is anyone's guess.
The production gap—11% of pilots reaching deployment—suggests a shakeout ahead, perhaps sooner than many founders anticipate. IDC predicts that by 2027, agentic automation will enhance capabilities in more than 40% of enterprise apps, with G2000 companies seeing agent use increase tenfold and token and API loads rise a thousandfold. Gartner expects cross-app multi-agent ecosystems by 2029 and autonomous resolution of 80% of common customer service issues by the same year. But those timelines assume governance, observability, and reliability challenges get solved in the near term—a big assumption given current state.
The Skeptics Weigh In
Not everyone is convinced the industry isn't getting ahead of itself. Andrej Karpathy, the AI researcher formerly at OpenAI, called current agents "slop" in October 2025, predicting it will take a decade to work through the issues. That's not a fringe view. The Camunda survey's finding that 73% of leaders see a gap between vision and reality suggests Karpathy's caution resonates well beyond the technical community. There's a growing sense that while the technology is advancing, the organizational and operational challenges—change management, workflow redesign, employee training—are proving stickier than anticipated.
For founders entering this market now, the question is less whether agentic AI will matter—the analyst forecasts and platform investments make that clear enough—and more where defensible value will actually accrue. Customer service, developer productivity, and healthcare back-office operations show early traction, but large SaaS vendors are already productizing agents in those areas. The hyperscalers' decision to embed agent builders and governance into their core stacks raises the bar considerably for infrastructure startups hoping to differentiate on security, observability, or total cost of ownership.
The YC batch composition may reflect a bet that application-layer startups can still carve out niches before incumbents close the gaps—that there's a window, however brief, to build something valuable before the platforms swallow the opportunity. It's a familiar Silicon Valley wager, one that's paid off spectacularly in some cycles and failed miserably in others.
But with only 11% of pilots reaching production, regulatory obligations tightening across major markets, and security questions still unresolved, the next 18 months will test whether nearly half a cohort betting on agents was prescient or premature. The technical substrate is in place, the capital is flowing, and the enterprise demand appears real. What remains uncertain is whether the operational and governance challenges can be solved fast enough to justify the valuations and expectations now baked into the market.
Time, as it tends to, will tell.
