Agent invents nonexistent SDK methods
You're asking an agent to use a third-party SDK you're unfamiliar with. Twice it has called methods that don't exist on the client (it's confidently inventing a plausible-sounding API), and twice it 'fixed' the error by inventing a different nonexistent method. How do you diagnose what's happening and re-steer it so the next attempt is grounded in the real API?
Implement
ground_calls_against_api(api_methods: list[str], called_methods: list[str]) → list[str]Examples
in
[["send_message","list_channels","close"],["sendMessage","close","fetchHistory"]]out["sendMessage=send_message","fetchHistory=missing"]in
[["connect","disconnect"],["connect","disconnect"]]out[]in
[[],["anything"]]out["anything=missing"]What a strong answer looks like
Treat the AI’s output as a draft to verify, not an answer to trust. Name the specific flaw and the input that triggers it, say how you’d catch it (tests, edge cases, reading critically), and how you’d re-prompt or decompose to get it right.
0:00 of about 15 min
Vibe & agentic: describe the solution in plain language (or narrate it) and the coach grades your approach.
Which questions mattered is sealed until you submit. Telling you now would just be handing over the edge cases.
Run or narrate your approach, then ask the coach.