Komodor has published a survey on the use of AI agents in engineering organisations. The findings suggest governance and control remain central concerns as teams expand agent-led operations.
Based on responses from more than 1,000 engineering, IT and technology leaders, the survey found that 53.6% cited building safe guardrails as a leading challenge. A further 50.4% pointed to governance and security issues, while 45.9% identified standardisation and accuracy among their top concerns.
At the same time, the research suggests AI agents are already being used at meaningful scale within engineering teams. Nearly 84% of respondents said they run fleets of 10 or more operational agents, indicating that for many organisations the question is no longer whether to deploy such tools, but how to manage them in production.
Restrictions remain common. The survey found that 75.7% of respondents still impose limits on autonomous operations, including read-only tasks, human approval requirements for changes, or constraints that keep automated actions away from critical workloads.
Scaling concerns
The results point to a gap between adoption and oversight. As engineering organisations give AI agents a role in production environments, many appear reluctant to let those systems act freely without clearer rules on access, execution and accountability.
The survey also found that 42.7% of respondents see insufficient visibility into agent permissions as a main problem. This suggests that control over what agents are allowed to do remains unsettled in a sizeable share of organisations.
Standardisation emerged as another pressure point. Nearly half of respondents ranked it among their top three challenges, reflecting concern over whether agents can deliver consistent, accurate results across different workflows and infrastructure environments.
Use cases are broadening across operations and software delivery. Routine task automation was cited by 49.1% of respondents, support ticket triage by 47.1%, self-remediation and AI SRE by 44.2%, and production incident investigation by 43.2%.
These figures suggest organisations are not limiting AI agents to peripheral tasks. Instead, they are increasingly applying them to operational work that affects production systems, incident response and service reliability.
Cost visibility
Spending is also rising. More than two-thirds of respondents, or 68.4%, said their organisations spend at least USD $10,000 a month on AI agents, including 38.4% spending USD $50,000 or more and 12% spending at least USD $250,000.
Yet visibility into those costs remains uneven. While 5.8% said they could not say how much their organisations spend, 25.8% ranked inadequate cost visibility, budgeting or spending limits among their top three challenges.
That combination of rising use and imperfect tracking may prove significant for engineering leaders trying to justify spending or prevent unexpected increases as fleets expand. The data suggests financial oversight has not kept pace with deployment in some organisations.
Another finding points to the effect of AI-assisted software development on operations teams. As coding agents contribute to faster development cycles, SRE, DevOps and platform engineering teams may face greater operational pressure from more deployments, more incidents to investigate and more infrastructure changes to manage.
CI/CD pipelines were cited by 36.8% of respondents as an area where AI agents are being used, adding to evidence that such tools are moving deeper into software delivery processes rather than remaining confined to experimentation.
Ben Ofiri, Co-Founder and Chief Executive Officer at Komodor, said the issues resemble earlier scaling challenges in cloud-native operations, but with a more consequential layer of autonomy. “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,” he said.
He added that AI agents differ in their ability to decide and act with less direct human input. “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,” Ofiri said.
The respondent base was weighted towards larger engineering organisations. More than 45% held Director-level roles or above, while nearly 85% worked in organisations with at least 100 engineers, including 28.5% in organisations with 500 or more.
Overall, the survey paints a picture of engineering teams extending AI agents into core operational work while still trying to put in place the guardrails, oversight and cost discipline needed to manage them safely.