Emerging tech & Deep tech • 1 day ago • Neha Jamwal
The first wave of enterprise generative AI was largely about adding intelligence to software that already existed. A CRM gained a copilot. A developer platform added code generation. An employee could ask a knowledge assistant to summarize documents or find information. The underlying enterprise architecture remained largely intact because the AI was still operating as a feature sitting on top of an application.
AI agents change that equation.
An agent does not simply generate an answer and hand control back to a person. It can interpret a goal, decide which tools it needs, retrieve information, interact with applications, execute a sequence of actions and adjust its approach based on what happens next. That makes the agent less like another software feature and more like a new participant in the enterprise technology environment.
And once software can act, rather than simply respond, infrastructure has to evolve around it.
The industry is consequently moving toward something bigger than an agent-building toolkit. Enterprises are beginning to assemble the foundations needed to run, control, secure and govern populations of AI agents. IBM, Microsoft and other major technology providers are already introducing concepts such as agent control planes, dedicated agent identities, lifecycle management and centralized governance.
That is the more interesting story behind the current agentic AI wave: AI agents are becoming an infrastructure problem.
From AI Features to Autonomous Participants
Traditional enterprise software generally has predictable boundaries. An application has a defined identity, an administrator controls its permissions, developers determine how it behaves, and security teams can map its activity back to known systems and users. Agents introduce a more fluid model.
An enterprise agent may operate across a CRM, ERP, ticketing platform, data warehouse and collaboration environment during a single workflow. It may select different tools depending on the situation. It may operate on behalf of an employee, another application or an automated process. In some architectures, agents may even interact with other agents. That changes the fundamental question enterprises need to ask.
It is no longer enough to know which application has access to what. Organizations increasingly need to know which agents exist, who owns them, what they are allowed to do, which tools they can invoke, what data they can reach and which human or business process ultimately remains accountable for their actions.
Microsoft’s emerging agent identity architecture reflects this distinction by treating agents as identities that require their own authentication, authorization and lifecycle controls rather than simply inheriting broad application or user credentials. This is a subtle but important architectural shift. Once an agent can take consequential actions, its identity becomes infrastructure.
The Emergence of an Agent Control Plane
The next challenge is operational.
An enterprise might begin with five agents. Then different business units create their own. Vendors introduce embedded agents into SaaS products. Developers build specialized agents for internal workflows. Employees create agents through low-code platforms. Very quickly, the organization has an agent estate rather than a handful of AI experiments. That estate needs something resembling the control planes enterprises already use for cloud infrastructure and distributed applications.
A mature agent control layer will increasingly need to provide:
- Discovery: What agents exist across the organization?
- Identity: Who or what does each agent represent?
- Ownership: Which team or individual is accountable for it?
- Authorization: Which systems, data and tools can it access?
- Lifecycle management: When should it be created, modified, suspended or retired?
- Observability: What decisions and actions has it taken?
- Policy enforcement: What actions require approval or additional controls?
- Governance: Does its behavior comply with organizational and regulatory requirements?
IBM’s agentic control-plane approach, for example, is explicitly designed around centralized visibility, governance and operational management of agents rather than simply giving enterprises another environment in which to build them. That distinction matters. Building agents is becoming easier; operating them responsibly at scale is becoming a harder engineering problem.
Identity Becomes the New Battleground
Perhaps the clearest sign that agents are becoming infrastructure is the attention now being paid to non-human identity. Traditional service accounts and API credentials were designed around relatively stable applications. Autonomous agents can be considerably more dynamic. They may be created for particular workflows, operate with delegated authority and interact with multiple systems during a single task.
Microsoft’s guidance describes this as a fundamentally different identity problem and recommends treating agents as first-class principals with appropriately scoped permissions. Its security guidance similarly emphasizes least privilege, explicit roles and tightly controlled tool access.
IBM has also begun introducing dedicated agent identity capabilities, underscoring the same architectural direction: organizations need to distinguish an agent’s actions from those of the person who initiated a workflow.
This could ultimately make agent identity management a new branch of enterprise IAM. The implications extend beyond authentication. Security teams will need to understand agent-to-agent interactions, delegated permissions, temporary access, tool authorization and the chain of responsibility behind an automated action.
Data Is Infrastructure for Agents Too
There is another layer that can easily get overlooked: agents are only as useful as the enterprise information they can safely access.
A conventional AI assistant can provide value from a relatively contained knowledge base. An autonomous agent executing business processes needs something more complicated. It needs current information, structured access to business systems, appropriate context and clear boundaries around what it is permitted to retrieve or modify. That puts pressure on data architecture.
Enterprises will increasingly need governed interfaces between agents and:
- Enterprise applications
- Databases and data warehouses
- APIs and business services
- Document repositories
- Knowledge bases
- Operational systems
- Real-time data streams
IBM’s current enterprise AI architecture, for example, connects agent orchestration with an AI-ready data foundation and hybrid-cloud operations, reflecting how tightly these layers are becoming intertwined.
The implication is significant: agent readiness may become another dimension of enterprise data readiness. Poorly governed data does not become safer simply because an AI agent is accessing it. In fact, autonomous access can amplify the consequences of weak permissions, stale information or fragmented governance.
The Infrastructure Stack Is Expanding
The emerging enterprise agent stack is therefore broader than a model and an orchestration framework. It is beginning to resemble a layered infrastructure:
Foundation models → Agent runtime → Tools and APIs → Identity and authorization → Data access → Orchestration → Observability → Governance and policy
Different vendors will package these capabilities differently, and the boundaries between layers will continue to shift. But the architectural direction is becoming clearer. The winning enterprise architectures will not necessarily be the ones with the greatest number of agents. They will be the ones that can deploy agents without losing control of the systems those agents touch.
That changes the role of enterprise architecture teams. Instead of evaluating AI solely as another application capability, architects increasingly need to consider agents as workloads with their own infrastructure requirements.
What CIOs and CTOs Should Be Thinking About
The immediate question for technology leaders is not whether every business process should become agentic. It is whether the organization has the foundations required to safely support agents when adoption accelerates.
That means addressing several questions early:
- Can the organization maintain a reliable inventory of deployed agents?
- Does every agent have a clearly defined owner?
- Can permissions be scoped to the minimum required actions?
- Can security teams distinguish human actions from agent actions?
- Can an agent be immediately suspended when its behavior becomes problematic?
- Can the organization reconstruct why an important action occurred?
- Are data-access policies enforced independently of the agent’s own reasoning?
- Can agents from different vendors operate within a common governance model?
These questions sound operational rather than futuristic. That is precisely why they matter. The infrastructure challenge of agentic AI is arriving before many organizations have finished establishing their governance models for generative AI.
The Real Shift Is From Intelligence to Agency
The enterprise AI conversation has spent years focused on intelligence: better models, better reasoning, better answers and better copilots. The next phase is about agency.
Once AI systems can act across enterprise environments, the architecture surrounding them becomes just as important as the intelligence inside them. Identity, permissions, data access, orchestration, observability and governance stop being supporting capabilities and become fundamental components of the AI system itself.
That is why the movement from AI features to infrastructure is more consequential than another generation of AI assistants.
The enterprises that treat agents as isolated application features may accumulate dozens of disconnected autonomous systems. Those that treat agents as a new infrastructure layer can begin building something more durable: an enterprise environment in which autonomous software can operate with defined identities, bounded authority, observable behavior and enforceable policies.
The agent era, in other words, may not be defined by how many agents an enterprise deploys. It may be defined by how well the enterprise can run them.
Key Takeaways
- AI agents are evolving from application features into an infrastructure layer, changing how enterprises think about architecture, security and operations.
- Agent identity is becoming a core enterprise security concern, as autonomous systems need their own authentication, permissions and accountability.
- Enterprises will need an agent control plane to discover, govern, monitor and manage growing populations of AI agents.
- Data architecture is becoming critical to agentic AI, with agents requiring governed, real-time access to enterprise systems, APIs and knowledge.
- Observability and governance will become as important as agent intelligence, particularly as agents begin taking actions rather than simply generating recommendations.
- The emerging agent stack extends well beyond foundation models, encompassing runtimes, tools, APIs, identity, data access, orchestration, observability and policy.
- The enterprise challenge is shifting from building agents to operating them safely at scale.