Jesse Stockall is Chief Architect, Flexera.
AI agents have officially moved from controlled pilots into everyday enterprise workflows. While this shift unlocks increased productivity, it risks creating a new layer of technology sprawl. Agents can access data, trigger actions and consume resources with limited human involvement. That makes them fundamentally different from much of the software enterprises are accustomed to managing.
Enterprises spent the first phase of generative AI asking, “What can we automate?” As agents move into production, architects now have to answer a harder set of questions: Who is accountable for them? What are they allowed to do? What should the business be willing to spend on them?
Negative consequences are not inherent to AI agents themselves, but they can emerge when organizations deploy agents faster than they establish the visibility and accountability needed to manage them. With key provisions of the EU AI Act having taken effect in August 2026, governance expectations around AI are also becoming increasingly concrete.
Beyond determining whether an agent can be connected, organizations must understand how they will identify, govern, observe and ultimately retire each agent. Every AI agent needs an owner because ownership provides the operating model for enforcing standards, governing costs, accounting for actions and scaling agents responsibly within enterprise architecture.
Here are a few key steps for building sound agentic architecture:
Define an owner.
Autonomy makes accountability essential. An owner is more than the person who built the agent: they are accountable for its purpose, compliance with enterprise standards, access to systems and data, authority to act, outputs, costs and retirement.
Ownership should follow the agent throughout its lifecycle and remain tied to an accountable person or team, just as the agent itself is tied to a distinct identity. Ninety-five percent of enterprises worldwide now have at least one company-funded agent-enabled workflow in production, with the average organization running roughly 11 across different functional areas. As that number grows, informal ownership will become increasingly difficult to sustain. If nobody can identify who is responsible for an agent after deployment, the organization does not truly control that agent.
Every cost, resource, license and action should be visible and attributable to an accountable user, whether human or autonomous.
Know what agents exist where.
Enterprises cannot assign accountability for agents they do not know exist, making a complete view of where agents are deployed and who is responsible for them essential. This challenge will only grow as agent adoption accelerates. Gartner forecasted that the average global Fortune 500 company could have more than 150,000 AI agents in use by 2028, up from fewer than 15 in 2025.
Foundational questions all architects should be able to answer are:
• What agents are running across the enterprise?
• Who created or authorized them?
• Which data sources do they touch?
• What permissions have they been granted?
• How much are they consuming?
AI activity can span models, SaaS applications, infrastructure, agents and employees. Every autonomous action must be traceable to its originating agent and accountable owner, with an audit trail and a clear path to rollback.
Give every agent boundaries, not just permissions.
Multi-agent systems, in which one agent delegates work to specialized subagents, have become a real architectural pattern. Companies such as Anthropic and OpenAI both have systems that support multi-agent architectures where a “manager” agent coordinates work across multiple specialized agents.
Traditional access controls answer an important question: “What can this system reach?” Agentic architecture also must answer another one: “What is this agent actually allowed to do once it gets there?”
Agents need enforceable boundaries beyond basic access control: limits on the actions they may execute autonomously, human approval thresholds for higher-risk actions, cost and budget thresholds, and shutdown mechanisms when behavior falls outside defined standards or expectations.
Agent architecture must therefore govern more than access. It must set standards for how and when agents may act, make their decisions about how and when to act, ensure actions are traceable to an accountable owner, and enforce limits on what those actions may cost.
Include cost as part of agent governance.
Cost should not sit outside agentic architecture as a reporting exercise conducted after deployment. Instead, it should remain a constraint an agent is designed to operate within.
An agent that delivers $100,000 of business value may justify substantial consumption, while an agent performing a low-value administrative task may not. Organizations therefore need cost attribution by agent and use case, visibility into consumption and enforceable budget thresholds. Because one workflow can consume models, tokens, infrastructure, SaaS resources and licenses across the technology stack, those costs must be assigned to an accountable human or non-human user rather than left in disconnected reports.
For architects, the goal should be to make an agent’s economic activity as visible and attributable as its technical activity, with both tied to an accountable owner. AI optimization should maximize return rather than simply minimize spend. Mission-critical applications and agents may warrant different spending limits based on their business importance, but those limits should be explicit, measurable and governed.
Treat agents like enterprise assets from day one.
The lesson from cloud and SaaS is that governance becomes much harder once adoption is already widespread. Enterprises have an opportunity now to avoid repeating that pattern with AI agents.
Agents should enter the technology estate with ownership, visibility, standards, controls and economic accountability already attached. Before an agent enters production, you should be able to answer four questions:
1. Who owns it?
2. What is it allowed to do?
3. How will its actions be traced and reviewed?
4. What is it worth spending on, and where will that cost be assigned?
The enterprises that scale agents successfully will not necessarily be the ones that deploy the most. They will be the ones whose architectures make every agent, resource, cost and action visible, attributable and controllable from the start.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?

