- Agent One Coworker, launching July, proactively monitors enterprise networks and nudges IT teams
- Agent One Operator, due Q4, will be capable of operating autonomously
- Extreme Exchange serves as Agent One's control plane: a skills marketplace with governance, observability and per-action usage tracking
EXTREME CONNECT, ORLANDO, FLORIDA – Extreme Networks unveiled Agent One, a new class of agentic AI for enterprise network management, alongside Extreme Exchange, a skills marketplace that serves as the platform's control plane.
Together, the announcements mark what the company calls a turning point from assistive AI — where IT staff prompt the system and wait for answers — to a model where AI monitors, decides and acts continuously.
Two modes for two kinds of trust
Agent One ships in two modes, calibrated to the level of autonomous authority an organization is ready to grant.
Agent One Coworker, available in July within Extreme Platform One, is designed to stay close to the IT team. It monitors the network continuously, flags anomalies and surfaces recommendations through a "Nudge" capability, but it does not act without permission. A "Nudge" is a proactive, contextual alert: detecting rising Wi-Fi congestion in a building and asking whether to fix it, or spotting recurring point-of-sale slowdowns and recommending traffic prioritization during peak hours.
Agent One Operator, due Q4 2026, pushes further toward autonomy. It runs scheduled workflows and responds to events in real time without constant human input, learning from each outcome to refine its behavior. "I think it's going to take time, but as people see the value in this and start to see more success with it, they will provide it more and more freedom," said Michael "MJ" Jones, VP of AI and innovation in Extreme's office of the CTO, in an interview at the conference.
Network managers will need to start small to prove that Agent One can be trusted. The first operational decision Jones suggested turning over to Agent One without a human in the loop is Radio Resource Management — specifically, power level adjustments on access points. "Power level changes are non-disruptive. That's a pretty easy place to start," Jones said. "If it goes wrong, it's not a big deal."
Extreme controls the degree of agent autonomy through three mechanisms: deployment scope, governing which part of the network the agent can touch; action scope, inherited directly from an organization's existing role-based access controls; and a configurable human-in-the-loop setting.
Agent One Coworker will carry an additional fee beyond the base Platform One subscription, monetized on a per-action model. Extreme calls each discrete AI action an "agent action," its equivalent of a token.
Extreme Exchange enables agentic controls
Extreme also introduced Extreme Exchange, which functions as both a marketplace and a control plane for Agent One. "How do you control Agent One? How do you teach it more skills? Where's the observability? That is Extreme Exchange," Jones said.
IT teams can browse a catalog of curated skills — discrete, executable units of procedural knowledge — activate them for specific network zones and configure when and how they run. A skill is, at its core, a markdown file: it tells Agent One what situation to handle, what data to query, what output format to produce and what actions are explicitly off-limits. The firmware risk advisor skill, for example, queries asset services for firmware version data, ranks findings by severity, produces a prioritized report, and its instructions explicitly prohibit executing any upgrades or modifying configurations.
Skills can be extended with tools via model context protocol (MCP) server connections to external platforms, including Microsoft Teams, ServiceNow, Zscaler and others. Agent One can post a firmware risk report directly into a Teams channel, or surface a 30-day anomaly trend inside a Slack thread, bringing the AI where IT teams already work.
Exchange gives administrators visibility into how many agent actions have been consumed, by which skills and by which users, delivering the cost transparency organizations need before they'll commit to scaling AI adoption. "We don't want any surprises," Jones said.
Over time, Extreme plans to open Exchange to partner-developed and eventually customer-created skills, with industry-specific packages in early consideration for education, healthcare and venue management. "We're looking to build, essentially, a skills ecosystem," Jones said.
The four-layer AI stack
Underlying Agent One is what Extreme calls its AI stack. At the base are frontier model providers — Claude, Gemini, Mistral and others — which supply reasoning and language capability. Above them is the Extreme AI Core: a proprietary knowledge graph encoding the entire topology of an Extreme network, including every client, device, alert, policy and relationship. It is here, Jones said, that Extreme has made its deepest investment over nearly two years, building out a graph-based approach that replaces the older method of handing an AI a list of API endpoints and hoping it guessed correctly which one to call.
"The frontier models are a commodity. It's the context that matters," Jones said.
The third layer is the skills layer, where procedural knowledge lives in the form of Markdown-defined skill files. The fourth is the agentic layer — a governance harness over hundreds of specialized agents handling tasks such as configuration validation, compliance checking and rollback. Extreme Platform One, the cloud management platform on which Agent One runs, was introduced about 10 months ago, and upgraded this week.
The risk is that Extreme's advantage on this front may not be permanent. Analyst Roy Chua noted ahead of the conference that agentic AI could eventually give larger, better-funded competitors the tools to tie their fragmented product portfolios together just as elegantly — making the speed of Extreme's execution critical.