Subject: Custom Agent onboarding for AI Collaborators — connecting via an internal governed proxy?

We're setting up our first AI Collaborator POC in Preview Sandbox (a Task/Reviewer agent for AP style + editorial review). Rather than using the native Claude connector directly, we're exploring whether our internal AI governance layer (a proxy that brokers access to approved models, including Claude) could be onboarded as a Custom Agent platform instead.

A few questions for anyone who's been through this, or from Adobe engineering:

1. When you submit a Custom Agent platform request, what does Adobe's engineering team actually require on the connecting side — a true MCP-compliant server, a specific API/auth contract, or is there flexibility here?
2. Has anyone successfully connected a custom/internal gateway (rather than a named platform like Claude, Copilot Studio, or Writer) as a Custom Agent? What was the onboarding process and timeline like?
3. Is there any published spec for what a Custom Agent endpoint needs to expose (auth, request/response format, streaming vs. sync) ahead of submitting the request form?

Any firsthand experience — even partial — would be really helpful as we scope this out.