I don't. I write a lightweight API proxy, usually as a CLI app, that takes requests from the agent, adds secrets and makes the request, and then passes the response transparently to the agent. The agent can use APIs without ever seeing secrets or tokens.

This approach doesn't scale and won't work in the cloud without a lot more effort, but that's not been a problem for me yet.