Full-stack vendor platforms
The deepest integrations come from vendors who control the hardware and telemetry pipeline end to end. These agents are built into the router, switch, and wireless operating systems themselves, which gives them access to signal a third-party tool doesn't get.
Marvis Virtual Network Assistant
HPE Juniper Networks
One of the earliest and most mature entries in this space. Marvis sits on top of Juniper's Mist AI telemetry and answers plain-language questions about wired, wireless, and SD-WAN health. "Marvis Actions" goes further and proactively remediates common issues: bad cables, missing VLANs, DHCP and 802.1X failures. "Marvis Minis" continuously simulate client connections in the background so problems surface before a real user hits them.
Cisco AI Assistant for Networking
Cisco Systems
A conversational layer across Catalyst Center and the Meraki dashboard, built on retrieval-augmented generation against Cisco's own documentation and telemetry, with filters aimed at catching hallucinated answers. The more interesting pieces are narrower: AI-RRM schedules radio channel changes during actual quiet hours instead of reacting to every blip, and an AI-driven packet analyzer can trigger its own captures and hand back a root cause instead of a raw pcap.
AVA (Autonomous Virtual Assist)
Arista Networks
Built on Arista's Network Data Lake, AVA's core trick is learning what "normal" looks like across a fleet so it can narrow in on the small residue of anomalous behavior rather than every deviation. That same approach has been extended into AI data center fabrics, watching ECN marks, PFC pause frames, and buffer utilization on GPU clusters. In the campus and SD-WAN side, "Ask AVA" translates zero-trust segmentation policy into plain language.
Extreme AI Expert
Extreme Networks
Part of Extreme Platform ONE, positioned as a multimodal conversational agent spanning both networking and security tasks instead of staying in one lane. Newer to market than Marvis or AVA, so the feature set is worth watching mature before taking the launch materials at face value.
Observability & overlay platforms
Vendor-agnostic tools that sit above a mixed-vendor environment, pulling in telemetry through open APIs and correlating across gear the vendor platforms above can't see into.
FlowAI & Intelligent Agents
Itential
The interesting design choice here is architectural: the AI never gets to actually touch the network. It proposes a workflow, and a separate deterministic orchestration layer enforces approvals, schedules the change into a maintenance window, executes it, validates the result, and rolls back automatically if something fails. It's one of the clearer answers to "how do you let a language model near production without losing control."
Selector Copilot
Selector AI
A multi-domain observability platform with a "network language model" front end, ingesting SNMP, gRPC, NetFlow, BGP, and syslog data. Engineers query across all of it in plain language instead of a proprietary query syntax.
Edwin AI
LogicMonitor
A narrower, ITOps-focused agent aimed squarely at alert fatigue: consolidating and reasoning across infrastructure alarms so a human sees one correlated incident instead of forty related pages.
AI Control Tower
ServiceNow
Aimed at telecom operators rather than enterprise networks: a control plane concept meant to connect business intent, risk, and network execution across OSS, BSS, and CRM systems. More of an emerging architecture than a single deployable product today.
Fabrix.ai Network Observability
Fabrix.ai
A vendor-agnostic observability layer, not a single-purpose agent: it ingests network, application, and infrastructure telemetry through more than a thousand pre-built connectors (Cisco, Juniper, Arista, Prometheus, Datadog, and more), correlates events across domains to find root cause, and can trigger closed-loop remediation through Ansible, Python, or Cisco NSO playbooks.
Security-centric agents
Agents built specifically for threat detection and policy enforcement, where the tolerance for error is lower and the models involved are often purpose-trained rather than a general-purpose LLM.
Precision AI / Strata Copilot
Palo Alto Networks
Powers Palo Alto's Cloud-Delivered Security Services with purpose-built security models rather than a general-purpose LLM, aimed at automated threat detection, zero-trust enforcement, and DNS/URL filtering at the speed those decisions need to happen.
FortiAI
Fortinet
Embedded in FortiAnalyzer, aimed at incident investigation and response: summarizing security events, estimating impact, and suggesting remediation so SOC analysts spend less time on triage and more on the incidents that matter.
Specialized agent catalogs
Rather than one monolithic assistant, some platforms take a modular approach: dozens of narrowly scoped agents, each built for one specific task, that can be combined for a given workflow.
Fabrix.ai is the clearest example of this pattern, with a public agent catalog spanning categories like NetOps, AIOps, SecOps, and Agent Ops. The exact number of agents in the catalog changes as they add more, so treat any specific count as a moving target rather than a fixed spec. A representative sample from the catalog is below.
I work at Fabrix.ai, so the six entries below aren't arm's length. See the full disclosure.
Network Change Risk Assessment Agent
Fabrix.ai
Scores the blast radius and rollback safety of a proposed network change before it goes out. Category: NetOps.
Network Config Compliance Agent
Fabrix.ai
Checks device configurations against vendor guidelines and internal best practices on an ongoing basis. Category: NetOps.
RCA for Networking Issues Agent
Fabrix.ai
Investigates the likely root cause of a networking issue within a given time window and suggests remediation. Category: NetOps.
Anomaly Detection Agent
Fabrix.ai
Learns baselines from metrics and logs and surfaces real-time deviations. Category: AIOps.
SecOps Incident Manager
Fabrix.ai
Automates incident handling: opens tickets, notifies the right people, and keeps the loop closed. Category: SecOps.
CVE Vulnerability Analyst Agent
Fabrix.ai
Rather than listing every CVE that touches installed software, cross-references known CVEs against the current network topology to estimate real exposure. Category: Cybersecurity.
Open source, community & research
The open source layer is where a lot of the experimentation is happening, from official vendor tooling to weekend projects to academic papers on how agents should route work to each other.
NetClaw
automateyournetwork (community)
Built on OpenClaw, one of the largest open source agent frameworks, NetClaw specializes it for network engineering: CLI/NETCONF skills, growing MCP integrations, and a Border/Member architecture delegating tasks to specialized sub-agents. One of the most visible, fastest-moving projects in this space, but still effectively a one-maintainer effort with no independent security review, and its own federation scheme (NCFED) is an individual IETF draft, not an adopted standard.