For the past few years, one of the biggest questions surrounding AI agents has been whether they could actually perform useful enterprise work.
That question has been increasingly answered. They can.
The more difficult questions have now shifted from technical capability to organizational deployment. Can enterprises trust agents to act on their behalf? How should they govern agents that interact with applications, data, and human employees, as well as other agents? Who is accountable when something goes wrong? And can companies capture enough economic value to justify the potentially rising costs?
These questions are at the heart of a recent Linux Foundation Research (LFR) report, “Enterprise Transitions to the Agentic Era,” written by Hilary Carter, Hillary Curran, and Anni Lai based on their interviews with 25 leaders, founders, enterprise operators, and advisors associated with Linux Foundation communities.
A 2024 McKinsey report nicely summarized this technological transition. We are moving from generative AI tools such as chatbots that answer questions and create content toward AI agents capable of executing complex, multistep workflows.
“In short, the technology is moving from thought to action.” And as AI moves from thought to action, the enterprise challenge changes as well.
The Agent as a New Coordinating Layer
The LFR report is organized around four pillars: technology strategy; governance and risk; business model shifts; and ROI and the agentic roadmap.
I found the first pillar particularly intriguing. Interviewees broadly agreed that agents are beginning to coordinate enterprise resources in new ways.
This raises an interesting question: could AI agents eventually play a role somewhat analogous to operating systems?
Operating systems have traditionally managed the hardware and software resources of computer platforms. More recently, cloud-native technologies have enabled enterprises to dynamically manage distributed computing resources and applications across cloud infrastructures.
Agentic systems may now be adding another layer of abstraction: coordinating applications, data, workflows, and increasingly skilled digital workers in pursuit of higher-level goals. Some study participants believed that enterprise agents are indeed becoming the kernel of a new generation of enterprise operating systems. As one put it, “just as the operating system abstracted hardware, the agentic layer is abstracting applications, data, and workflows.”
Others felt that it’s far too early to make such a claim. A few compared agents to schedulers rather than operating systems: agents don’t own the underlying resources but can prioritize access to them based on inferred intent. And others argued that enterprises still need to develop the equivalent of the orchestration frameworks that emerged during the transition to cloud-native computing.
The precise analogy remains unsettled. But the broader point is important. Agents don’t replace the fundamentals of enterprise computing or software engineering. They add a new execution and coordination layer on top of them.
And once AI agents begin coordinating enterprise resources and taking actions on their own, governance becomes essential.
From Retrospective Oversight to Real-Time Governance
The strongest point of agreement among the interviewees was the need for a governance layer between agents and the resources they now manage.
The report identified four major responsibilities for such a layer:
- Policy enforcement, to evaluate consequential actions before they occur;
- Identity and authorization, to establish who or what an agent is acting on behalf of;
- Metering and attribution, to track usage and costs; and
- Auditability, to maintain an immutable record of actions and their outcomes.
This highlights a fundamental difference between generative and agentic AI.
A chatbot typically generates information or a recommendation and leaves it to a human to decide what to do. An AI agent may itself take action — accessing a database, interacting with an application, initiating a transaction, or handing a task to another agent.
Governance therefore cannot rely primarily on discovering that something went wrong with the action taken by an AI agent.
The interviewees agreed that retrospective audits are insufficient when agents operate at machine speed. Policies and guardrails increasingly need to be enforced, — either at the runtime level or through the model’s own self-regulation, — before consequential actions are executed.
In other words, as AI moves from generating recommendations to taking actions, governance must move from retrospective oversight to real-time control.
Who Is Accountable for Digital Coworkers?
Governance also raises the difficult question of accountability for virtual employees.
On this point, the interviewees largely agreed: regardless of an agent’s degree of autonomy, responsibility remains with the organization deploying it and with the people who configure and authorize its actions. Blaming a failure on the AI itself is unlikely to satisfy regulators, customers, or the public.
But accountability becomes considerably more complicated in a multi-agent environment, — when an agent might delegate a task to another agent, which then accesses a service or calls still another agent. The report identifies accountability across such interacting agents as an unresolved frontier.
The traditional AI governance question has been: Where is the human in the loop?
The agentic era adds another: how should we govern handoffs between multiple agents? As organizations deploy hundreds or potentially thousands of autonomous agents, manually overseeing every interaction will clearly be impossible and require considerable automation.
Agentic AI Will Change Software Economics
The transition from software as a tool to software that increasingly performs work also has significant implications for enterprise business models.
The report found broad agreement that traditional per-seat software pricing will come under pressure. If agents perform work previously carried out by people, pricing software according to the number of human users becomes increasingly disconnected from the value the software delivers.
Also, most interviewees also rejected predictions that “Software-as-a-Service is dead.” They expect established SaaS companies to adapt rather than disappear.
Cost pressures may also lead enterprises toward hybrid portfolios of AI models. Smaller, often local open source models could handle high-volume routine tasks, while more expensive frontier models are reserved for problems requiring sophisticated reasoning.
The common thread is that when software increasingly performs work rather than simply helping people perform work, the economics of enterprise software must evolve as well.
Measure Outcomes, Not Agents
How, then, should enterprises determine whether their agentic AI investments are paying off?
The interviewees shared a straightforward principle: measure completed work and value delivered, not activity.
The number of agents deployed or tokens consumed tells us little about whether an organization is becoming more productive. Instead, companies should focus on metrics such as completed tasks, resolution rates, quality, time saved, and fully loaded cost per outcome compared with the previous human-centered process.
The report also offers a useful perspective on failed AI pilots.
Rather than viewing high pilot-failure rates as evidence that agentic AI isn’t working, interviewees generally regarded failure as a normal part of experimentation. But experiments should have clearly defined objectives and record their results so that failures produce organizational learning rather than simply additional expense.
Agentic AI may also diffuse more rapidly than cloud computing did. Agents run on cloud infrastructure that enterprises have already built and learned to trust. Open source can significantly accelerate development. And adoption is increasingly bottom-up, with employees already experimenting with AI tools whether or not their organizations have formally approved them.
An Enterprise Roadmap for the Agentic Era
When participants were asked about the biggest obstacles to adoption, trust and governance ranked first, followed by data and context, change management, and cost.
These barriers suggest a useful roadmap for enterprises.
Build trust and governance. Organizations need clear identities, permissions, policies, audit trails, runtime guardrails, and accountability before giving agents significant autonomy.
Invest in organizational knowledge and context. The report makes an especially important point: “an agent’s effectiveness is bounded by the context available to it.” Knowledge and context management should therefore be treated as core infrastructure rather than an afterthought.
Enterprises have spent decades accumulating knowledge across databases, applications, documents, APIs, and the minds of their employees. Agents will only be as useful as their ability to securely access and correctly interpret that organizational context.
Redesign work rather than merely insert agents into existing processes. The transition to agentic AI is as much an organizational transformation as a technological one.
Measure economic outcomes. Focus on the value of completed work rather than tokens consumed, numbers of agents deployed, or other activity measures.
Experiment aggressively but govern the experimentation. Bottom-up experimentation can accelerate learning, but organizations should convert ungoverned shadow AI into enterprise managed adoption and ensure that successful and unsuccessful pilots both contribute to institutional knowledge.
Capability Is No Longer the Gate
AI agent capability is not the gate is the most important message from the Linux Foundation Research report.
That doesn’t mean AI agents have solved every technical problem. Far from it. But the central enterprise challenge is increasingly shifting from making agents more capable to building the organizational infrastructure that enables them to act safely and productively.
That infrastructure includes identity, authorization, runtime guardrails, auditability, cost controls, high-quality data, organizational context, and clear human accountability. It also requires companies to rethink workflows, software economics, and how they measure the value of work performed by combinations of humans and digital coworkers.
Some important questions remain unsettled. We don’t yet know whether agents will become the kernel of a new enterprise-wide operating system, which technical standards will prevail, how dramatically agentic AI will reshape today’s software industry, or whether adoption will actually proceed faster than the cloud native transition. The report appropriately distinguishes such contested questions from areas where consensus is already emerging.
But the broad direction is becoming clearer.
Generative AI gave us enterprises systems that could increasingly understand and generate language, software, images, and other forms of content. Agentic AI is now giving those systems the ability to act. That changes the nature of the enterprise challenge.
The transition to the agentic era will ultimately depend less on how intelligent AI agents become than on how intelligently enterprises learn to govern, integrate, and work with them.
AI disclosure: I wrote the initial draft of this blog and used ChatGPT for editorial assistance. I remain responsible for the final content.