Why same-Wi-Fi AI agent discovery fails on larger home networks

Your phone and Mac can show the same Wi-Fi icon and still fail automatic discovery.

I ran into that exact problem while testing Hermes Mobile, the Android control surface I build for remote AI-agent sessions. The phone completed its “find computers” sweep, the Mac gateway was healthy, and the firewall was off. Yet the Mac never appeared.

The failure was not mystical networking. It was a subnet assumption.

The evidence that separated the app from the network

The useful debugging order is simple:

1. Confirm the Mac-side pairing service is listening on a phone-reachable interface.
2. Confirm its pairing document answers over the Mac's LAN address.
3. Confirm the agent gateway health endpoint answers over that same LAN address.
4. Compare the actual subnet mask with the address range the phone scans.

In this case, both services were healthy:

pair server: 0.0.0.0:8765
agent gateway: 0.0.0.0:8642
pair.json: HTTP 200
/health: HTTP 200, status=ok
macOS application firewall: disabled

That ruled out the common explanations: Hermes was open, the gateway was not bound only to localhost, and macOS was not rejecting the connection.

The decisive evidence was the network configuration. The Mac was on a network with mask 255.255.252.0, which is a /22. That one network spans four adjacent /24 ranges.

The mobile scanner, however, generated candidates by keeping only the first three IPv4 octets from the phone and trying hosts 1 through 254. In other words, it scanned one /24 even when the real Wi-Fi network was larger.

If the phone received an address in one quarter of the /22 and the Mac received an address in another, they were legitimately on the same Wi-Fi network but the scanner could never probe the Mac.

Why “same Wi-Fi” is not the same as “same /24”

Many discovery implementations quietly assume a home network looks like this:

phone: 192.168.1.42
Mac:   192.168.1.79
mask:  255.255.255.0

That assumption works until a router, mesh system, office network, or lab uses a broader subnet. On a /22, all of these addresses can belong to the same local network:

192.168.68.x
192.168.69.x
192.168.70.x
192.168.71.x

A scanner that only substitutes the final octet will miss three quarters of the valid address space.

This is why a progress bar can reach 100% without finding a healthy computer. The progress accurately represents the addresses selected by the app, but the selection itself is incomplete.

A better discovery contract

The durable fix is not to make the progress animation slower or add another retry. Discovery should derive its candidate range from the phone's IP address and subnet mask, then bound the work so it remains safe on a real device.

A robust contract should:

• calculate the network prefix instead of assuming /24;
• probe known saved hosts first;
• prefer a previously successful LAN address;
• probe explicit Tailscale or MagicDNS hosts when available;
• cap concurrency and timeouts to protect phone memory and battery;
• show the exact fallback path when automatic discovery cannot prove a match.

There is also a product lesson here: “Finding…” is not enough. A remote-control app should tell the user whether it is scanning LAN addresses, checking a saved Tailscale route, waiting for a pair server, or failing authentication. Honest state makes a networking problem diagnosable.

The manual path should remain first-class

Automatic discovery is convenience, not identity. A safe fallback is an explicit computer address paired with a short-lived setup flow. That avoids broad scanning and gives the user a deterministic route when the network topology is unusual.

Hermes Mobile supports local and remote connection paths because AI-agent control should not depend on one fragile assumption about a router. The goal is straightforward: follow an agent's work from your phone, send messages, and approve or deny risky tool calls while the actual agent and its credentials remain on the computer you operate.

I am sharing this failure because it is more useful than claiming that local discovery “just works.” It worked on /24 networks and missed a valid Mac on a larger subnet. The endpoints were healthy; the address-generation logic was incomplete.

If you want to try the Android control surface, install Hermes Mobile: AI Agent Leash on Google Play. It is a paid-upfront companion for controlling a compatible Hermes setup on your own computer—not an on-device chatbot.

The implementation and setup material are available in the public source repository.