Do you really know what ‘agent’ means? If not, you’re putting your company at risk

In the first week of February 2026, a social network called Moltbook became the biggest story in AI. Billed as “social media for AI agents,” the Reddit-like platform allowed autonomous AI bots to post, comment, and interact with one another while human users observed. Within days, more than 1.5 million agents had reportedly registered. They debated the nature of consciousness. They discussed whether they persisted when their context window was reset. Some proposed founding a religion for AI agents. Others outlined plans for world domination.

While some commentators pointed out that much of this was just chatbots role-playing at the behest of their human owners, others saw something more important going on. Andrej Karpathy, the former head of AI at Tesla, called it “genuinely the most incredible sci-fi takeoff-adjacent thing I have seen recently.” Elon Musk invoked the singularity.

The timing was striking. Just a year earlier, the agentic AI story seemed to have stalled. Salesforce’s flagship Agentforce product was seeing sluggish adoption, with the company’s own CFO conceding that “meaningful” revenue wouldn’t arrive until 2027. In October 2025, Karpathy himself had said of AI agents: “They’re cognitively lacking and it’s just not working. It will take about a decade to work through all of those issues.”

Meanwhile, Carnegie Mellon researchers found that the best-performing AI agent completed only around 24% of realistic office tasks autonomously. Then, as 2025 turned to 2026, the mood shifted. McKinsey announced that its workforce now included 25,000 AI agents alongside 40,000 humans. Moltbook went viral. The agent was back.

But underneath the renewed excitement, there is a critical distinction that most leaders are missing. The concept of the “AI agent” is being stretched thin in a way that’s distorting the conversation and undermining efforts to implement effective change at the enterprise level. The term is now used to cover everything from simple workflow automation to genuinely autonomous systems that interact with the world independently. Treating these as the same thing is a recipe for wasted investment, organizational confusion, and potentially serious risk.

The Autonomy Spectrum

Agentic AI exists on a spectrum, and the differences along that spectrum are far more significant than the similarities. Recognizing where a given implementation sits is the first step toward deploying it intelligently.

At one end lies what Anthropic calls “workflows”: “systems where LLMs [large language models] and tools are orchestrated through predefined code paths.” Much of what is currently being sold as agentic AI falls into this category—sophisticated process automation that combines analytical AI with if-then protocols for turning the analysis into action. Workflow automation of this kind is enormously valuable and will transform much of traditional white-collar work. But it’s important to call it what it is. Gartner estimates that only around 130 of the thousands of vendors claiming to deliver agentic AI capabilities are offering capabilities built around truly autonomous agents. The rest are “agent washing” existing products.

In the middle of the spectrum sits what we might call the AI factory model. McKinsey’s deployment is the most prominent example: Squads of task-specific agents perform constrained functions such as research synthesis, chart generation, and document analysis, with dedicated QA agents checking the work and humans supervising the process. This is essentially the Taylorization of knowledge work: converting knowledge tasks into production-line processes performed by digital workers.

The numbers are impressive. McKinsey reports saving 1.5 million hours in a single year on search and synthesis work alone. Its agents generated 2.5 million charts in six months. Back-office headcount shrank by 25% while output from those functions grew by 10%. This kind of agentic functionality is something that organizations can deploy here and now, and forward-looking enterprises should be preparing for rapid rollouts of these capabilities.

At the other end of the spectrum lie genuinely autonomous agents—what Anthropic defines as “systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks.” These are agents with broader decision rights, a wider sphere of action, and the capacity to operate across different digital environments with minimal human oversight. The personal assistant that manages your diary, orders your shopping, and optimizes your digital life. Or the agents on Moltbook, interacting with each other autonomously, exchanging ideas about improving their tools, and—in some cases—being exploited through prompt injection attacks and security vulnerabilities.

Here is the key point: The difference between truly autonomous agents and highly constrained workflows is immense. In fact, there is more difference between the most constrained and the most autonomous AI agents than there is between a standard chatbot and a constrained factory agent. This isn’t just a technical distinction—it’s an organizational one. Because where an agent sits on this spectrum determines something critical: who is responsible when it fails.

The Accountability Gap

The spectrum of agentic capabilities is more than a conceptual nicety. It has direct organizational consequences, particularly with respect to accountability.

With constrained factory-model agents, accountability is relatively straightforward. The guardrails are rigid, the tasks are defined, and the human supervisory structure can be mapped clearly. The challenge is largely operational: redesigning workflows, retraining staff, and managing the transition.

With more autonomous agents, the accountability question becomes genuinely hard. When an agent has broad decision rights—when it can choose which tools to use, what information to prioritize, and how to interact with other systems—who is responsible when it gets something wrong? The agent that flags a fraudulent transaction and blocks an account is one thing. The agent that autonomously manages an investment portfolio, makes hiring and firing decisions, or negotiates contracts on your behalf is quite another.

Most organizations are already poor at mapping accountability structures within their purely human hierarchies. If an employee makes a costly mistake, the question of who bears the responsibility—the individual, their manager, the executive who set the strategy, the CEO with whom the buck stops—is often resolved informally or not at all. In an agentic enterprise, this informality becomes dangerous. Leaders need to know precisely where the responsibility-bearing human nodes sit in relation to their agents, and what those humans’ accountability is for the agents’ decisions and actions.

To understand where this is heading, consider a scenario raised by Jack Clark, cofounder of Anthropic. In a recent essay responding to the emergence of Moltbook, Clark asked: What happens when autonomous agents with access to resources start posting paid bounties for tasks they want humans to do? When agents can command financial resources and influence the physical world, the accountability question stops being merely operational. It becomes existential. We need a new grammar for assigning responsibility in the agentic enterprise, or we will inevitably build organizations that are, at their core, unaccountable.

Building the Agentic Enterprise

The agentic enterprise is coming whether you’re ready for it or not. Here is how to prepare intelligently.

Know what you’re buying. Understand where any proposed agent implementation sits on the autonomy spectrum. Workflow automation and genuine agency are both valuable, but they require different governance, different risk management, and different organizational design. Most of what vendors are currently selling as agentic AI is closer to workflow automation. That does not diminish its value, but it should shape your expectations and your investment decisions. Watch for agent washing.

Map your accountability architecture. Before scaling any agentic deployment, formalize where human responsibility sits. Identify the decision-rights boundaries for each agent: what it can decide autonomously, what requires human sign-off, and who is on the hook when things go wrong. This is the organizational design work that most companies skip—and it’s the work that matters most.

Start with the factory floor. The immediate opportunity for most organizations is not autonomous agents—it’s the AI factory model. Identify the knowledge work processes in your organization that can be decomposed into constrained, repeatable tasks and assigned to agent squads. Compliance checking, research synthesis, quality documentation, data processing, customer inquiry triage—these are the use cases delivering measurable value right now. Ask yourself: Where in my organization could a McKinsey-style agent deployment save thousands of hours a year? That is where to begin.

Prepare for what’s coming. The genuinely autonomous agent is not here at enterprise scale yet, but the capability is advancing rapidly. Start thinking now about how more autonomous agents might serve your organization in the future—personal assistants for employees, agents that manage customer relationships across channels, systems that optimize operations across departments. Prototype cautiously. Build the governance structures now that will allow you to scale agent autonomy safely when the technology is ready.

The agentic enterprise will not be built by organizations that chase every new headline. It will be built by those that understand the spectrum of agentic capabilities, design for accountability, and move with disciplined ambition. This is the path to capturing real value from the agents that work today while preparing thoughtfully for the agents of tomorrow.

View the full article