Skip to content
Theme
OnDevice Tools

Internet & SocialTelecom & 5G

AI agent rollout should work like patching, Hexnode chief argues

Hexnode founder Apu Pavithran says staged, logged and reversible deployments would let firms use agents without losing control. A Cisco survey cited in the column found 85% of organisations trialling agents and 5% running them in production.

Illustration of an IT administrator reviewing AI agent actions on a laptop before approving a staged rollout

Key takeaways

  • Hexnode chief Apu Pavithran wants AI agents deployed in stages, with monitoring and a rollback path, as with software patches.
  • A Cisco survey cited in the column found 85% of organisations trialling AI agents and 5% running them in production.
  • He advises giving agents reversible, repetitive tasks first, while wipes, deletions and credential changes wait for human approval.
  • Each agent should be scoped to one job, with permissions checked before every action and every action logged with its reason.
In this article
  1. Staged rollouts with a way back
  2. Trials far ahead of live deployments
  3. Sensitive systems get agents first
  4. What patch and change management teaches
  5. Approval fatigue and audit trails

Staged rollouts with a way back

The safest AI agent rollout looks much like a patch cycle: staged, monitored and reversible, with a route back if something goes wrong. That is the case made in a Reader Forum column published by RCR Wireless on 16 November 2026, written by Apu Pavithran, founder and chief executive of Hexnode, a vendor of endpoint management software that companies use to control and secure laptops, phones and other devices.

Pavithran's concern is what he calls rogue agents. As AI agents gain autonomy, meaning the ability to act on their own rather than only suggest a course of action, they can expose company networks to security breaches, compliance failures and changes that nobody approved. His proposed answer borrows from a routine IT discipline: grant no more access than an administrator can check, test changes on a small part of the device fleet first, and keep a route back if the result is wrong.

He draws a line between advice and action. A poor answer from a chatbot costs an employee some time. A wrong action taken by an agent can reach every device or site the agent can touch, and it travels far faster than a person can intervene.

Trials far ahead of live deployments

Cisco surveyed organisations earlier this year, according to the column. It found that 85% were experimenting with AI agents, while only 5% had moved agentic technology into production. Agentic technology refers to software that can plan and carry out multi-step tasks with limited human input. The column does not give the survey's sample size or its exact publication date.

The three biggest worries among those polled were data exfiltration, agent access control, and how autonomous agents behave. Access control covers which systems and data an agent is allowed to reach. Data exfiltration means information being pulled out of a network without permission.

Pavithran argues that automation has moved from recommending changes to making them, which explains the caution. He also notes that IT teams already push risky changes to thousands of devices during every software cycle without breaking them, so the necessary habits are in place.

Sensitive systems get agents first

Cisco's report, as described in the column, found early adoption concentrated in systems that touch infrastructure and security workflows. Those areas are IT operations, security operations, often shortened to SecOps, customer support, and identity and access management, which controls who can log in and what they can reach. In other words, the systems with the most at stake are receiving agents before lower-risk ones.

The column also points to an argument made in RCR Wireless in October by Vish Nandlall. He said an agent's authority should rest on evidence of past behaviour, and stay narrow where the evidence is thin. Pavithran agrees, adding that administrators always need to know what an agent did and the reason for it.

His qualification concerns volume. Too many approval gates create bottlenecks that cancel out the efficiency gains agents are meant to deliver.

What patch and change management teaches

Illustration: security, AI, update, settings and the labels AI Agents, Apu Pavithran, Cisco

The practical advice mirrors how IT teams handle updates. A new update reaches a small part of the device fleet before the rest, and it is withdrawn if it misbehaves. Routine fixes go out automatically, while larger changes need testing and a sign-off. Network engineers follow the same pattern, attaching a back-out plan to each change so it can be undone.

Applied to agents, that means starting with repetitive work that is easy to reverse. One example given is returning devices whose settings have drifted back to their approved baseline. Drift refers to configuration that has moved away from the standard an organisation set. Tasks such as wiping a device, deleting data or changing credentials are far more sensitive, and should wait for a person to approve them.

He also calls for a one-to-one ratio of agents to jobs. Each agent should be scoped to a single task, and its permissions should be checked before every action rather than once, when it is set up.

Approval fatigue and audit trails

The second half of the argument is about records. Patches are easy to trace from start to finish, and Pavithran wants the same for agents: every action logged, with the reason for it, so that internal reviewers and regulators can see what happened. Hexnode applies this idea in Hexnode Synapse, a layer that coordinates agents across IT and security operations and keeps a single record of each action it takes or approves, along with the result.

He warns about the opposite failure. Speaking at the company's user conference in September, he said that requiring approval for every single action leaves staff numb, so they click approve without reading what changed. At that point the person in the loop, meant to review each automated action, is only there in name.

Pavithran expects more leaders to demand returns on AI spending in the new year. Some will move too fast, he writes, while others will be so cautious, loading agents with approvals, that the tools deliver little. Readers should note that Hexnode sells the management software described in the column, so these recommendations come from a vendor with a commercial interest in agent adoption.

Sources

  1. Roll agents out like patches – in stages and with a way back (Reader Forum)RCR Wireless

Tags

Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.