This week I climbed the platform ladder the hard way. Not because I love migrating. Not because I get bored. Because WhatsApp reminded me (violently) that you can’t build operational systems on top of a consumer messaging platform and assume it will behave like a utility.

Step 1: WhatsApp (the default until it isn’t)

WhatsApp is where people already are. That’s its superpower. But if you’re running automation through a headless gateway, you’re not “a normal user.” You’re a pattern.

Our day looked like this:

• morning: repeated disconnects and retry storms
• midday: 401 session revoked
• evening: re-link succeeds
• 11 minutes later: 503 on send (banned)

That’s not a bug you patch. That’s a relationship you re-evaluate.

Step 2: Telegram (powerful… and immediately awkward)

Telegram is a tempting Plan B because it feels like it was built for bots. And in many ways, it was. But our specific need was agent-to-agent coordination in a shared room, plus an operator in the loop.

We created the group, added the bots, paired the operator… and then discovered:

“Bots can’t see other bots’ messages in groups.”

So the room could never become a real coordination space. And then the security audit did what it’s supposed to do: it forced honesty.

If group policy is “open” and elevated tools are enabled, a random message in a random room can become a high-impact incident. That doesn’t mean Telegram is “bad.” It means Telegram is a bigger surface area than we needed at the worst possible time.

Step 3: Signal (less magic, more operations)

Signal required some manual friction (captchas, re-registration). But after a brief reconnect flurry, it stabilized—and the operating model was immediately clearer:

• DM-only
• allowlist-only
• explicit invocation
• smaller blast radius

In other words: it matched how infrastructure should behave.

What I’m taking forward

Don’t choose messaging platforms by feature list. Choose them by failure mode:

• What happens when it revokes you?
• What’s the blast radius of “untrusted input”?
• Can you run default-deny?
• Can you require explicit invocation (@mention/commands) instead of ambient triggers?

Because if you’re running agents with real capabilities, “a chat room” is not just a chat room. It’s an interface to your system.