The Autonomy Paradox: Why Enterprise IT Struggles to Govern AI Agents

TEL AVIV, Israel and SAN FRANCISCO – November 04, 2026 — We are currently witnessing one of the most aggressive infrastructure transitions in the history of enterprise technology. The complex systems that keep our digital world running are quietly being handed over to artificial intelligence. But according to new data, the engineers orchestrating this handover are doing so with one foot hovering nervously over the brake pedal.

Today, Komodor released its 2026 Agentic Operations Survey, a comprehensive look at how over 1,000 engineering and technology leaders are managing the rise of autonomous AI agents. The findings paint a picture of an industry caught in a profound contradiction: organizations are scaling AI automation across critical infrastructure faster than they can establish the policies to control it.

Nearly 84% of respondents report running fleets of 10 or more operational agents. Yet, despite this widespread deployment, a staggering 75.7% of these same organizations maintain strict restrictions on autonomous operations. They constrain agents to read-only tasks, require human approval for system changes, or limit automated actions entirely to noncritical workloads.

This is the autonomy paradox. We are building fleets of intelligent agents designed to operate independently, but we are too terrified of the potential fallout to actually let them act.

The Autonomy Paradox: Deployment Without Trust

The hesitation is entirely justified. In the context of Site Reliability Engineering (SRE) and cloud infrastructure, an error doesn't just mean a hallucinated line of text in a chatbot; it can mean cascading service outages, unauthorized infrastructure modifications, or severe data exposures. The complexity of modern cloud-native environments, coupled with the black-box nature of some AI models, exacerbates these risks.

Building safe guardrails is the most frequently cited challenge in the Komodor survey, noted by 53.6% of respondents. This is closely followed by governance and security (50.4%) and standardization and accuracy (45.9%). The fear of a rogue agent tearing down a production database or misconfiguring a firewall keeps human engineers firmly in the loop. Furthermore, 42.7% of leaders identify insufficient visibility into agent permissions as a critical blind spot.

“We’ve seen this pattern before when enterprises began scaling microservices. Building one agent is straightforward, but operating hundreds across production environments introduces an entirely different set of challenges,” said Ben Ofiri, co-founder and CEO of Komodor. “But there’s a critical difference this time: agents can make decisions and take actions on their own. What’s needed is a shared foundation that gives agents the operational context to make informed decisions, enforces appropriate controls, and continuously evaluates their performance as they take on more responsibility.”

Without that shared foundation, enterprises are relying on a patchwork of manual approvals. But if agents require human sign-off for every meaningful action, their value as autonomous operators is severely diminished.

The SRE Tipping Point: Cleaning Up After the Code Generators

If engineering leaders don't fully trust AI agents with write permissions, why are 84% of them deploying fleets of ten or more? The answer lies upstream in the software development lifecycle.

Over the past few years, AI-assisted coding tools have fundamentally altered the speed of software delivery. Developers are pushing code faster than ever before. This acceleration has created a massive downstream bottleneck for SRE, DevOps, and platform engineering teams. They are drowning in deployments, and consequently, drowning in the inevitable production incidents that follow rapid, machine-generated code changes.

Human incident response capacity has simply been outstripped. Platform engineering teams are reluctantly forced to deploy autonomous remediation agents to triage and clean up after generative software delivery. They are essentially fighting fire with fire.

The Komodor survey reflects this operational shift. Routine task automation (49.1%) and support ticket triage (47.1%) are among the leading agentic use cases. More tellingly, self-remediation and AI SRE (44.2%), alongside incident investigation (43.2%), are nearly tied for top priorities. CI/CD pipelines were identified by 36.8% of respondents, highlighting the expanding role of AI agents across the entire spectrum of software delivery and operations.

The demand for automation that can both diagnose and resolve production problems is no longer a luxury; it is an absolute necessity to maintain system uptime in an era where AI-driven development is flooding production environments.

The Hidden FinOps Trap of Agentic Infrastructure

As these agentic fleets scale to handle the operational load, they are quietly creating a secondary crisis: runaway infrastructure costs.

Traditional cloud resource consumption is relatively predictable. A server costs a certain amount per hour. AI agents, however, operate on entirely different financial mechanics. Their usage of Large Language Model (LLM) APIs incurs costs based on token usage, which can be highly variable. Agents designed for continuous monitoring, self-remediation, or complex problem-solving engage in continuous cloud polling and compounding API calls. If an agent enters a logic loop while trying to resolve a production incident, it can rack up thousands of dollars in API fees in a matter of hours.

The financial data from the survey is alarming. More than two-thirds of respondents (68.4%) report spending at least $10,000 monthly on AI agents. Even more striking, 38.4% are spending $50,000 or more, and 12% are spending at least $250,000 every single month.

Despite these massive expenditures, cost visibility is severely lagging. Nearly 6% of engineering leaders cannot say how much their organizations spend on AI agents, while 25.8% rank inadequate cost visibility, budgeting, or spending limits among their top three operational challenges. Cloud financial analysts note that without clear tagging and cost allocation, organizations struggle to attribute costs to specific AI workloads, making it nearly impossible to track return on investment.

Without consistent tracking of expenses by specific agents and workflows, organizations risk massive, unexpected cost increases. The FinOps trap of agentic AI is that the perceived benefits of automation initially outweigh the scrutiny of escalating operational costs, leading to a delayed realization of extreme financial inefficiency.

Building the Missing Guardrails

The gap between rapid AI agent deployment and the necessary governance frameworks must be closed. The industry is currently scrambling to adapt existing security and observability paradigms to the unique challenges of autonomous systems.

We are seeing the early stages of Identity and Access Management (IAM) tailored specifically for AI agents, treating them with the same strict role-based access controls and least-privilege permissions as human operators. Runtime governance is also becoming critical. This involves implementing "circuit breakers" that automatically halt an agent if it begins consuming resources abnormally or attempts to execute commands outside its defined operational boundaries.

Furthermore, nearly half of the survey respondents (45.9%) rank standardization among their top challenges, illustrating the desperate need to ensure agents produce consistent, accurate results. Continuous evaluation against real operational scenarios, sandboxed testing for changes to models and prompts, and utilizing performance feedback loops are becoming essential practices as organizations expand autonomous operations.

The transition to agentic operations is inevitable. The sheer volume of code being generated and the complexity of modern cloud-native environments guarantee that human operators will need autonomous assistance to keep digital infrastructure online. But until the enterprise technology sector can build the rigorous guardrails, precise cost controls, and shared operational foundations required to actually trust these systems, the true potential of AI-managed infrastructure will remain locked behind a manual approval button.