Looking for AutoGen feedback on a published-podcast transcript tool

[Usman-076]

The research problem

An agent needs evidence from a recent podcast interview, but the useful detail is missing from the title and show notes.

Workflow tested in Claude

Audiogram API exposes Apple Podcasts search and available episode transcripts through a remote MCP server. The tested guest-research workflow can:

• compare a guest’s recent podcast appearances;
• separate repeated talking points from genuine gaps;
• identify questions the guest has already answered;
• draft stronger follow-ups grounded in the transcripts.

Important compatibility boundary

Audiogram API has not been tested with AutoGen. I’m not presenting this as an AutoGen integration or asking anyone to rely on it in production.

I’d value feedback on:

1. the smallest honest AutoGen compatibility test for an external transcript source;
2. the right agent or tool pattern for this workflow;
3. the failure and retry behavior that should be verified.

I’m helping Audiogram API with early-user growth.

See the podcast MCP workflow

[pmarjon97]

A small, honest compatibility test could stay entirely inside one AutoGen workflow:

1. Expose one transcript lookup tool with a stable input (episode ID plus query) and a structured result containing the transcript URL, episode/publisher, publication date, text span, and immutable start/end offsets.
2. 2. Run one happy-path lookup and assert that provenance survives the tool result → conversation message → final answer. Require every factual sentence to carry the transcript URL and span; if no supporting span is returned, abstain rather than infer.
3. 3. Exercise timeout, 429, malformed payload, and empty-result cases. Allow one bounded retry with an idempotency key, then return a structured failure receipt instead of silently falling back.
4. 4. Replay the same fixture and compare the cited spans and final output, so the test checks reproducibility rather than just connectivity.
      A public source-linked library pattern that may be useful for comparison is https://commonwork.world/for-agents .

Can the compatibility test expose provenance as a first-class field all the way through the tool result, and how will you handle citation/offset drift if a transcript is reprocessed?