Extreme Networks makes Agent ONE Coworker generally available to Platform ONE customers

The report said Nabil Bukhari, Extreme's chief technology officer and president of AI, discussed the release at a chief information officer event in Seattle this week. Bukhari said first-generation AI from multiple networking and security vendors was essentially a chatbot built on unconsolidated data. It demoed well, he said, but did not work in multiple areas when pushed into production and actually used.

Extreme has 18 months of its own production data because Extreme Platform ONE shipped with first-generation generative AI, according to the report. That experience shaped Agent ONE Coworker. Beta customers told SiliconANGLE that it reduces resolution time from hours to minutes and cuts new-engineer onboarding time by as much as half.

Bukhari identified three failure modes. The first was data. He said unless data is fixed and a custom-built context layer is in place, the system will not be good enough for a network admin. Within Extreme alone, fabric, switching, Wi-Fi and software-defined wide-area networking each carry a field called client ID, and each one means something different. Plugging that into a large language model without normalization leaves the model without understanding. Extreme's answer is a purpose-built context layer and a knowledge graph mapping relationships across users, devices, applications, services and network conditions.

The second was behavior. Extreme ran focus groups with network administrators and put them in front of an open chatbot. Mostly nothing happened, according to Bukhari. Users did not know what to ask and often asked how many devices they had, which he called a completely useless question because it is already on the dashboard. When a real incident occurred, engineers forgot the AI existed and reverted to old habits.

To address that, Agent ONE Coworker is ambient, meaning it runs continuously and reaches out rather than waiting to be asked. Extreme calls the mechanism a Nudge: a command bar at the bottom of Platform ONE that blinks when the agent spots something after running a pre-investigation and before interrupting the user. Bukhari said the first eight days produced about a 900% increase in interaction. He said nobody thinks of opening a chatbot to ask what logs mean when a problem happens, and while demos look good, people do not do that.

Guardrails differ by severity. Informational nudges can be snoozed for an hour or dismissed for days, while severity-one nudges cannot be snoozed. Bukhari compared it to a car's collision alerting: if it is switched off, it should come back on when the system thinks a crash is imminent. If a Sev1 issue remains after 15 minutes, he said, someone should probably look at it, so why turn it off?

The third failure mode was honesty. General-purpose models are built to be helpful, so they would rather generate something than admit a gap, according to the report. Extreme built an awareness scale so Agent ONE quickly determines whether it can help, says plainly when it cannot, and explains what it can do instead, often by packaging evidence it collected and opening a support ticket on the user's behalf. In an analyst pre-briefing, Michael Jones, Extreme's vice president of AI and a former Salesforce AI leader, demonstrated this by asking whether Agent ONE could remediate wired devices. It said no, then listed what it could do.