Summary
Storyboards should be the primary developer interface for building an AdCP agent — prescriptive before you code, evaluative as you build, and a reference for what passing looks like.
Based on analysis of 9 real production conversations with Addie where developers tried to build/test/connect agents, plus two expert reviews.
Key findings from real conversations
- The Welcome to adcp Discussions! #1 interaction is "test this URL" — developers paste a URL and want instant feedback
- Platform type creates confusion — Addie asks "What platform type?" which developers can't answer correctly
- v2->v3 is the real blocker — every agent test fails on v3. Step-by-step execution would surface this at step 1
- Developers want to understand what is being tested before running it
- Buyer vs seller confusion — developers say "buyer agent" when they mean seller MCP server
- No seller quickstart exists in any language
The flow: discover, recommend, run
- Discover — paste URL, connect, discover tools (fast, non-destructive)
- Recommend — "Your agent exposes X tools. These storyboards apply. Start with media_buy_seller?"
- Confirm — developer picks or accepts recommendation
- Run step-by-step — one step at a time, pass/fail with reference comparison
Implementation phases
| Phase |
What |
Where |
| 1 |
Seller quickstart in JS, Python, Go |
adcp-client, adcp-client-python, adcp-go |
| 2 |
Step-by-step runner endpoint + UI |
this repo |
| 3 |
Tool-based storyboard resolution API |
this repo |
| 4 |
Storyboard map UI (replace flat picker) |
this repo |
| 5 |
Error humanizer for validation failures |
this repo |
| 6 |
Addie storyboard tools (recommend, detail, run, progress) |
this repo |
| 7 |
Result persistence + regression notifications |
this repo |
| 8 |
Remove org gate for storyboard testing |
this repo |
Summary
Storyboards should be the primary developer interface for building an AdCP agent — prescriptive before you code, evaluative as you build, and a reference for what passing looks like.
Based on analysis of 9 real production conversations with Addie where developers tried to build/test/connect agents, plus two expert reviews.
Key findings from real conversations
The flow: discover, recommend, run
Implementation phases