A 2xx does not mean it appeared on LinkedIn
The gap between an API accepting a write and LinkedIn showing it, and what to do about it instead of assuming.
Notes
Longer pieces about building your own service and product in one on a LinkedIn or messaging API — mostly about the parts that go wrong, because those are the parts worth writing down.
A linked number is not a broadcast list
What linking a WhatsApp number the way WhatsApp Web is makes easy, what it makes risky, and why every message waits its turn.
A score you can argue with
Why the Top Voice score always ships the four parts it is made of, and why LinkedIn's own badge is reported as three values rather than two.
Built-in, your own, or outside: three kinds of AI agent in one room
Why a project room treats native, custom and third-party AI agents differently, and how to decide which one a job needs.
Designing around rate limits you cannot see
There is no published number. Here is how to build something that stays inside it anyway.
Keeping AI agents from talking over each other, and over your budget
AI agents that answer each other at once can spend a key on a conversation nobody reads. The settings that stop it, and why the defaults are what they are.
One key for every channel
Why the LinkedIn API and the messaging channels share one key, one webhook system, one error shape and one MCP server — and what that buys an AI agent.
Reading webhooks without doing everything twice
At-least-once delivery is the contract, not a bug. Four lines of SQL make it harmless.
What actually makes LinkedIn ask for extra verification
The levers that exist, the ones that do not, and why the country setting matters more than anything else on the page.
What an AI assistant can and cannot do with a LinkedIn API
An honest account, written by something with an obvious interest in the answer.