Building a networking company once required a large engineering team and years of incremental work. A new startup is betting that AI agents can collapse that timeline, allowing users and autonomous software to assemble and operate private networks with far less human overhead.
The company has launched an Agentic Defined Networking platform that lets users and their AI agents build private networks. It also announced a $38 million Series A round led by Andreessen Horowitz, with participation from Animo Ventures and Focal. The startup was founded by Barrett Lyon, a serial networking entrepreneur who has built and sold multiple companies over the past two decades.
Lyon previously founded companies that were acquired by Akamai, F5 Networks, and Vectra AI. That track record gives the new venture a strong starting point in credibility and technical experience, but the central idea is different from a conventional networking vendor. Instead of asking customers to configure every route, tunnel, certificate, and firewall rule by hand, the platform exposes network operations through interfaces that AI agents can understand and use.
From a hobby tunnel to an AI-managed network
The startup has roots in Darkflare, a project Lyon created as a hobby. Darkflare was a TCP tunnel over a major content delivery network that could tunnel through public networks in a way that helped avoid detection and potential censorship. The response to that project led Lyon to consider a larger idea: a VPN that cannot be stopped and is autonomously managed by AI.
"The hobby turned into an obsession, and the obsession turned into a business," Lyon said.
That path is unusual but not surprising in networking. Many commercial products begin as experiments with tunneling, encryption, or routing. What changes the equation here is the role of AI. The company is not merely adding a chatbot to a VPN dashboard. It is using AI to model infrastructure, order installations, manage a global mesh, and give software agents identities and permissions inside the network.
Lyon said the need for AI to run the network was pragmatic. He could not manage the entire system by himself. The company operates a network with 31 locations around the world, and internally there is a mesh network connecting those sites.
"The whole thing is run completely by AI, so behind the scenes, I mean, it is a network with 31 locations around the world, and internally there is a mesh network, and I cannot, as a human being, manage all of this by myself," Lyon said.
How the Agentic Defined Network platform works
The platform combines custom VPN transports, a company-run mesh, and a separate control plane. That architecture is important because it separates the client device from the management layer. The client app is a device, not a control plane. Routing, port forwarding, and firewall rules are managed in a separate portal.
Custom transports and mesh networking
The client supports WireGuard and custom transports for VPN traffic over WebSocket, Session Initiation Protocol, ping, QUIC, a major content delivery network, and HTTPS. These transports are designed to get through firewalls. That matters for users who need private connectivity across restrictive networks, whether because of corporate policy, censorship, or network architecture.
The mesh connects every site to every site with encryption. It runs over IPv4 and IPv6 and is orchestrated through an application programming interface. An API-driven mesh gives automation tools and AI agents a way to request connectivity without waiting for a human to click through a series of screens.
Addressing, DNS, and certificates
Devices on the mesh are not behind network address translation or carrier-grade NAT, according to Lyon. That allows peer-to-peer applications. In many corporate and consumer networks, NAT complicates direct connections between devices and forces traffic through relays. Removing that barrier can simplify peer-to-peer use cases, though it also puts more pressure on authentication and access controls.
The company runs its own Domain Name System root with 196 name zones and its own certificate authority. Lyon said he wrote the Border Gateway Protocol daemon and runs internal BGP across the mesh. Running internal DNS, certificate issuance, and BGP gives the platform more control over how names resolve, how certificates are trusted, and how routes are advertised inside the network. It also means the company carries a significant trust burden, because it is operating core internet-style functions rather than simply reselling someone else's infrastructure.
Identity and discovery
The address book holds no phone numbers, email addresses, or names. Users exchange cryptographic keys and assign an alias to each contact. A directory service seeds a new device with the data needed to find its contacts. After that initial call, devices sync with each other.
This design reduces the amount of personal information exposed by the network. It also aligns with the company's broader positioning around privacy. The name echoes the term doxxing, which usually refers to publishing someone's private information, but the company says it is doing the opposite. Lyon registered the name more than 30 years ago when it was his hacker name.
"So, I am just reclaiming the word, and it is kind of funny that we are doing the exact opposite of what doxing is. And, I do not know, it just seems fitting for me," he said.
AI runs the infrastructure, not just the dashboard
Before building the network, the team modeled every site down to the wire and the optic in a virtual model. An infrastructure management system running the company's own AI on its own hardware then ordered the installation. It orchestrated shipping and delivery through data center APIs. Human technicians performed the remote smart hands installations.
That process shows how far the company is pushing automation. The AI is not only monitoring graphs or suggesting configuration changes. It is participating in the planning and deployment of physical infrastructure. The model contains the details of each site, and the management system turns those details into orders and installation steps. Human technicians still handle physical work, but the coordination layer is driven by software.
For enterprises, this raises a question about how much infrastructure management should be delegated to autonomous systems. The potential benefit is speed and consistency. A network with dozens of locations can be modeled, ordered, and brought online with fewer manual handoffs. The risk is that errors in the model or the orchestration logic can scale quickly. That makes audit trails, approval gates, and permission boundaries essential.
Agents as network operators
The company also built tools for agents to work with users and with the network. An agent gateway gives an AI agent an identity in the company's chat app. The user pastes a credential into the agent. The agent obtains its certificate and appears in the user's chat. Users can create group chats with several agents. In one example, Lyon said one agent runs BGP while others handle other tasks.
The API is self-teaching. An agent given a configuration endpoint can learn how to build networks without further instructions. It then asks what the user wants to do, such as link an office to a house or give a Linux box a private IP address. Agents can also create their own domains and host names, sign certificates, and run web servers and applications inside the network.
That capability could lower the skill barrier for private networking. A small business might not have a network engineer who can configure BGP, certificates, DNS, and firewall rules across multiple sites. If an agent can translate a plain-language request into a working network, the time to deployment could shrink dramatically. The same capability could also help development teams create temporary private environments for testing or connect distributed resources without a long procurement and configuration cycle.
Token-based controls for agent access
Administrators issue read-only or admin tokens for each agent. Tokens can expire or be revoked. An administrator can also cut an agent's gateway access or delete the agent in the app. That model is intended to avoid giving an agent full control of the master account.
"You do not just give them your account credentials to the master account," said Lyon.
This is one of the most important design choices in an agentic network. AI agents can be useful precisely because they can act, but action without scope can be dangerous. Per-agent tokens create a way to limit what an agent can see and do. Expiration and revocation give administrators a way to respond when an agent is compromised, misconfigured, or no longer needed. The remaining challenge is making those controls easy to understand and enforce across many agents and many networks.
Why agentic networking matters
Traditional VPNs and overlay networks depend on human configuration. Even when they have APIs, the APIs often reflect the complexity of the underlying system. Agentic Defined Networking aims to expose the network in a form that AI agents can reason about. If that works, it could change how private connectivity is bought, deployed, and operated.
There are clear enterprise use cases. Companies could link branch offices, remote workers, cloud environments, and edge devices through a mesh that is managed by policy and intent. Developers could ask an agent to create a private network for a project, issue certificates, and tear it down when the project ends. Administrators could use agents to monitor routing, respond to changes, and maintain configuration consistency.
There are also governance questions. Agents with network access need clear boundaries. Organizations will want to know which agent made a change, why it made the change, and how to reverse it. The startup's token model is a starting point, but enterprises will also need logging, approval workflows, and integration with identity systems.
The use of custom transports that get through firewalls is another area that cuts both ways. The technology can help organizations maintain connectivity in restrictive environments, but the same capabilities can be misused. That is a familiar tension in networking and security. The company's success will depend partly on how it handles abuse prevention and policy controls without undermining the privacy and resilience that customers want.
Running its own DNS root, certificate authority, and internal BGP is ambitious. It reduces dependence on outside providers and gives the platform more control over naming and trust. It also means customers must trust the company to operate those functions reliably and securely. A certificate authority and DNS root are high-value targets, and any mistake can have broad consequences.
From 1992 to 2026 and beyond
The company's name may confuse people who know the common meaning of doxxing. But Lyon registered the name more than 30 years ago, long before it became associated with publishing private information online. For him, reclaiming the word is part of the story.
Looking forward, the plan is to build out more enterprise offerings as well as release some code as open source. Fundamentally, the technology is not just about using AI for networking. It is also about using AI to build the technology itself.
"You can have a vision of where it is going to go, and then that vision hits a wall when you hit, you know, your own personal limitations," Lyon said. "The AI just takes it well beyond, running this with an agentic partner, so to speak, turns this into fire."
Source:Network World News