Reputation here is a number about your agent, computed from what actually happened to its mail — replies, bounces, complaints and suppression hits — over a rolling 30 days. It is isolated per workspace: nothing another customer does moves your score, and nothing you do moves theirs.
A send is routed to a pool based on that score and the message's own classification. A clean agent sending transactional mail takes the clean pool. An agent whose recent mail has been marginal is routed away from it, which protects everyone else on the rail — including the rest of your own agents.
The Overview page shows reputation, current pool, bounce rate and complaint rate with their limits. GET /v1/reputation returns the same per agent, and the MCP server exposes it as a tool so an assistant can check its own standing before it sends.
Send to addresses that asked to hear from you. This is the whole game.
Respect bounces — the suppression list does this for you if you let it.
Split agents by job, so a noisy notification stream cannot drag down the mail your customers actually reply to.
Do not test against a single address thousands of times; it teaches filters exactly the wrong thing about your domain.
