On a winter day early this year, a startup called Vanka AI demonstrated something that felt unsettling in the way the future often does when it arrives unannounced. Feed the system a brief business concept, the pitch went, and within minutes it would spin out a complete startup scaffold: brand strategy, go-to-market plan, visual identity, UI specifications, even a GitHub repository with working code. No human touched the keyboard during the demo. Discord served as the control plane, where executive agents—each playing CEO, CTO, or CMO—debated product direction and made decisions among themselves.
Compelling, yes. Ready for prime time? Absolutely not. But the demonstration signaled a shift that's now unfolding across the enterprise software landscape, faster than many expected: the move from AI that assists to AI that operates.
By late March, that vision had funding to match the ambition. Sycamore—led by former Atlassian CTO Sri Viswanath—closed a $65 million seed round to build what it calls a "trusted agent OS for the enterprise." The round, led by Coatue and Lightspeed, ranks among the largest seed raises on record for an infrastructure startup. That same week, Microsoft open-sourced its Agent Governance Toolkit, tackling the runtime security challenges that arise when autonomous systems start moving money and data without human checkpoints.
The timing wasn't coincidental. After years of chatbot experiments and productivity copilots, observers note the industry is pivoting hard toward autonomous execution. The question is no longer whether AI can draft an email or summarize a meeting. It's whether AI can run the meeting, execute the strategy, and close the quarter.
The Platform Play
Microsoft's Agent 365, expected to roll out as part of its new E7 bundle, represents perhaps the clearest articulation of where the big platforms are headed. It's not a single agent but an enterprise control hub—a registry, analytics layer, and orchestration system for the hundreds of agents that Microsoft anticipates will proliferate inside large organizations. The company's messaging around "the Frontier Firm" partnership with Harvard positions agents not as tools but as a new category of workforce. Which is either visionary or a bit much, depending on your tolerance for vendor hyperbole.
Salesforce made a similar bet earlier, launching Agentforce in 2024 and positioning Slackbot as "the front door to the agentic enterprise" when it went generally available to Business and Enterprise customers in early 2026. AWS brought multi-agent collaboration to general availability last year, showcasing a case study with Syngenta where autonomous agents coordinate across data aggregation, recommendation, and conversation—delivering crop yield improvements the company pegged near 5%. Google followed with Workspace Studio, a no-code agent builder that entered general availability late last year.
IBM's watsonx Orchestrate, meanwhile, expanded its agent catalog and observability capabilities in 2025, aligning governance with the Model Context Protocol—Anthropic's open standard for connecting agents to tools and data. The protocol, announced in November 2024, has become something of a Rosetta Stone for the ecosystem. Though security researchers discovered and patched critical vulnerabilities in its Git server implementation as recently as late last year, which should give pause to anyone rolling this out at scale.
Gartner forecast last August that 40% of enterprise applications will embed task-specific AI agents by the end of 2026, up from less than 5% in 2025. By some measures, we're on track. Deloitte's TMT Predictions report suggested that up to 75% of companies may invest in agentic AI this year. KPMG's Global Tech Report found that 88% of companies are already investing, with high performers reporting a 4.5× average ROI compared to the 2× industry average. Those numbers have the ring of analyst optimism, though the directional trend seems undeniable.
What a Company OS Actually Does
The term "Company OS" or "Agent OS" has become shorthand for platforms that model an organization's departments, processes, and resources, then deploy governed agents to automate cross-functional workflows. Think less virtual assistant, more autonomous business layer.
Sierra, the customer experience platform co-founded by Bret Taylor, articulated the vision when it announced Agent OS 2.0: "a single agent across every channel that learns from every interaction." The company had raised $350 million at a valuation reported around $10 billion. Its pitch is that agents should have persistent memory and context across customer touchpoints, replacing the fragmented bot experience with something that actually gets smarter over time. Whether it lives up to that promise remains to be seen—early enterprise deployments tend to reveal edge cases that demos don't.
Sycamore's approach, judging from the sparse public statements, emphasizes trust and security—building the governance layer first rather than bolting it on later. "Operating system for autonomous enterprise AI," Viswanath told press at the funding announcement, highlighting "trust, security and control" as foundational requirements. Which is probably the right instinct, given how quickly things can go wrong when autonomous systems have access to production databases and customer accounts.
Other startups are staking out adjacent territory. /dev/agents raised a $56 million seed round to build what it calls a cloud agent OS with ex-Google, Meta, and Stripe leadership. Airrived emerged from stealth with $6.1 million to unify agentic workflows across cybersecurity, IT, and operations. SupraOS positions itself around "verifiable enterprise execution"—governed autonomy rather than unchecked decision-making.
Even the open-source community is experimenting in the wild. WizardingCode published a tutorial for building "ARKA OS," a seven-department agent system with persistent knowledge and connections to more than a dozen external tools. Unwind AI released a framework designed to "replace the C-suite" with autonomous CEO, CMO, and CTO agents, complete with Telegram integration and oversight via the Model Context Protocol. That last bit reads like satire but apparently isn't.
Not all of these will survive contact with reality. The pattern, though, is clear: the infrastructure layer for autonomous business operations is being built right now, in public, with real capital behind it.
The Reality Check
Dell Technologies CTO John Roese warned earlier this year about "agent washing"—vendors slapping the label on anything with an API and calling it agentic. His prediction: 2026 would reveal which claims hold up under scrutiny. So far, he seems prescient.
Gartner has been more pointed. Across multiple reports published over the past year, the firm has maintained that more than 40% of agentic AI projects will be canceled by the end of 2027. The reasons: unclear value, escalating costs, and inadequate risk controls. A BCG and MIT Sloan report found strong intent signals—73% of "agentic AI leaders" expect competitive differentiation—but acknowledged that broader adoption remains early and skills gaps are real.
Security researchers, for their part, have demonstrated exactly why governance matters. Vulnerabilities in the LangChain and LangGraph frameworks—widely used for building agent systems—were disclosed earlier this year, exposing pathways for data exfiltration across multiple classes of enterprise data. The bugs have been patched, but they illustrated the risks of giving autonomous systems broad tool access without hardened runtime controls.
Microsoft's Agent Governance Toolkit offers one answer: a framework for implementing kill switches, concentric security rings, and saga orchestration patterns that allow agents to operate within defined boundaries. The fact that Microsoft felt compelled to open-source this before Agent 365 even launched suggests the company understands the stakes. Or perhaps they've learned from watching other vendors ship first and govern later.
Regulatory timelines add another layer of complexity. The EU AI Act's main provisions take effect in August 2026. Colorado's AI Act goes live on June 30, 2026. California's Transparency in Frontier AI Act has been in force since the start of the year. Each imposes disclosure, impact assessment, or incident reporting requirements that could apply to autonomous agent deployments, depending on how they're classified and what decisions they make.
The Zero-Employee Experiments
Which brings us back to the provocative question: Can you actually run a company with no humans?
Metavesco, a public company, announced in March that it was launching a "zero-human company initiative" as an internal business unit. The press release was short on operational detail and long on aspiration. Independent validation is scarce, and the announcement reads more like a proof of concept than a going concern.
More instructive, perhaps, is a Reddit post from a 23-year-old founder who documented 90 days of building what he called a "zero-human company." Posted in March, his lessons were candid: automation itself is straightforward; distribution and quality control loops are much harder. The operational work gets delegated to agents. The strategic oversight, customer acquisition, and trust-building still require humans in the loop. In other words, the boring parts can be automated. The parts that matter, not so much.
Vanka AI's demonstration showed what end-to-end scaffolding looks like—strategy, brand, code, infrastructure. But scaffolding isn't the same as a sustainable business. You can automate the production of a startup; whether it can acquire customers, adapt to market feedback, and survive first contact with competition is a different question entirely.
What Founders Should Actually Watch
The infrastructure is maturing faster than most anticipated. OpenAI's Operator model, available via API since last year, achieved a 38.1% success rate on OSWorld benchmarks at launch—a score that multiple vendors have since improved upon. Amazon's Bedrock platform supports multi-agent collaboration with supervisor-collaborator patterns proven in production at companies like Syngenta. Google's no-code builder lowers the barrier to experimentation.
Interoperability is improving, slowly. The Model Context Protocol and the Agent2Agent standard, backed by more than 100 companies under the Linux Foundation, are creating common substrate for cross-platform agent coordination. Microsoft's Agent 365 will likely introduce a unified registry and API, allowing enterprises to manage agents from multiple vendors through a single control plane. Whether vendors will actually adopt it in practice—or build walled gardens instead—is the more interesting question.
The talent and mindset shift, however, may be the harder part. IDC has forecast 1.3 billion agents by 2028—a number Microsoft has cited in its own materials. If even a fraction of that projection materializes, companies will need new operating models. Not just prompt engineers or AI trainers, but process architects who can redesign workflows for autonomous execution and governance specialists who can define agent behavior boundaries.
The companies getting this right aren't treating agents as a bolt-on productivity feature. They're rethinking org charts, decision rights, and accountability structures. They're asking what work requires human judgment and creativity versus what can be safely delegated to systems that operate around the clock, don't get tired, and scale horizontally without hiring freezes.
The zero-employee company remains more thought experiment than blueprint. But the zero-additional-employee company—the one that doubles revenue with the same headcount by layering in autonomous systems—is already here. Whether that's a feature or a warning depends on how carefully we build the governance, security, and oversight mechanisms that keep those systems aligned with human intent.
The race is on, certainly. The tooling exists. The capital is flowing. What remains to be seen is whether the industry can ship the necessary guardrails as quickly as it ships the capabilities. History suggests that's the harder problem to solve.
