A survey by engineering recruiting firm Howdy.com found that two-thirds of knowledge workers who use AI also use AI agents, and 16% have started building agents themselves. The survey did not ask whether employees built them without organizational approval, but Howdy.com founder Frank Licea said many employees are doing so without waiting for permission. He said some employees were asked to pause until guardrails and system guidance for safe development were in place. Workers want to automate tedious manual tasks, Licea said, and every organization has employees who are curious about technology.

About one in four agent users said they had disabled an agent because its output was inconsistent or it did not work properly from the start. Seven percent of knowledge workers said they had to stop an AI agent that had gone out of control. In recent months, Howdy.com employees have shared ways they automated manual tasks without waiting for IT approval. Technology staff hearing about these examples asked which API keys were used, whether encrypted data was accessed, and how the work affected SOC 2 compliance.

Luis Mesas, vice president of product engineering at software engineering services firm Sngular, said employees will keep building agents regardless of whether companies formally allow it. He said the barriers to development are falling while engineering teams are being asked to do more with fewer resources. The central management challenge, Mesas said, is not agent development itself but agents connecting to company systems and data without the controls applied to regular software.

Mesas said IT leaders need to understand how employee-built agents connect to internal systems and data. If an agent retrieves customer information from a customer relationship management (CRM) system, changes records, or moves information between systems, leaders should check the scope of its access and whether the employee needs those permissions. Mesas described this as the point where an agent that began as a personal productivity tool becomes a company-wide technology management issue.

Andy Sen, CTO of B2B commerce platform provider AppDirect, said employees should be encouraged to build agents within clear guardrails and policies. Shadow IT has always existed, Sen said, and shadow agents are its latest form. AppDirect gives employees a digital experimentation environment and developed the AI-powered coding service Devs.AI in response to employee demand. Sen said the service lets employees build and innovate freely while giving IT visibility into usage and costs, and confidence that company data is not leaking externally.

Sen said IT leaders should pay close attention to the finding that one in four employee-built agents is disabled. Failure and learning are acceptable, he said, but failures should not happen out of sight; leaders need to understand agent behavior before small issues become serious. He was skeptical of reports that agents had gone out of control, arguing that in most cases an agent followed its instructions, but the instructions were wrong. Responsibility, he said, lies with the developer rather than the agent.

Andrew Sales, principal methodologist at agile methods firm Scaled Agile, said the high rate of disabled agents is a signal for IT leaders to investigate the causes. He said agents built quickly outside management processes often lack boundaries around their tasks, the data they can access, and when human intervention is needed. AI governance should not focus only on preventing bad outcomes, Sales said; it should also create a path for turning good ideas from individual experiments into organizational capabilities. Identifying effective examples, standardizing them, ensuring they are safe, and responsibly expanding them across teams can help organizations find problematic agents while spreading successful approaches.

A survey by engineering recruiting firm Howdy.com found that two-thirds of knowledge workers who use AI also use AI agents, and 16% have started building agents themselves. The survey did not ask whether employees built them without organizational approval, but Howdy.com founder Frank Licea said many employees are doing so without waiting for permission. He said some employees were asked to pause until guardrails and system guidance for safe development were in place. Workers want to automate tedious manual tasks, Licea said, and every organization has employees who are curious about technology.

About one in four agent users said they had disabled an agent because its output was inconsistent or it did not work properly from the start. Seven percent of knowledge workers said they had to stop an AI agent that had gone out of control. In recent months, Howdy.com employees have shared ways they automated manual tasks without waiting for IT approval. Technology staff hearing about these examples asked which API keys were used, whether encrypted data was accessed, and how the work affected SOC 2 compliance.

Luis Mesas, vice president of product engineering at software engineering services firm Sngular, said employees will keep building agents regardless of whether companies formally allow it. He said the barriers to development are falling while engineering teams are being asked to do more with fewer resources. The central management challenge, Mesas said, is not agent development itself but agents connecting to company systems and data without the controls applied to regular software.

Mesas said IT leaders need to understand how employee-built agents connect to internal systems and data. If an agent retrieves customer information from a customer relationship management (CRM) system, changes records, or moves information between systems, leaders should check the scope of its access and whether the employee needs those permissions. Mesas described this as the point where an agent that began as a personal productivity tool becomes a company-wide technology management issue.

Andy Sen, CTO of B2B commerce platform provider AppDirect, said employees should be encouraged to build agents within clear guardrails and policies. Shadow IT has always existed, Sen said, and shadow agents are its latest form. AppDirect gives employees a digital experimentation environment and developed the AI-powered coding service Devs.AI in response to employee demand. Sen said the service lets employees build and innovate freely while giving IT visibility into usage and costs, and confidence that company data is not leaking externally.

Sen said IT leaders should pay close attention to the finding that one in four employee-built agents is disabled. Failure and learning are acceptable, he said, but failures should not happen out of sight; leaders need to understand agent behavior before small issues become serious. He was skeptical of reports that agents had gone out of control, arguing that in most cases an agent followed its instructions, but the instructions were wrong. Responsibility, he said, lies with the developer rather than the agent.

Andrew Sales, principal methodologist at agile methods firm Scaled Agile, said the high rate of disabled agents is a signal for IT leaders to investigate the causes. He said agents built quickly outside management processes often lack boundaries around their tasks, the data they can access, and when human intervention is needed. AI governance should not focus only on preventing bad outcomes, Sales said; it should also create a path for turning good ideas from individual experiments into organizational capabilities. Identifying effective examples, standardizing them, ensuring they are safe, and responsibly expanding them across teams can help organizations find problematic agents while spreading successful approaches.