Bob looks up the movie times
Reconstruction⚠ This is not a live capture: it is a condensed reconstruction, edited afterwards from the real transcript of two sessions, on September 12 and 14, 2026. The prompts are the ones from the actual sessions (in French), cut in places: the cinema's town is replaced with « […] » and a web address pasted with one prompt is left out. The timing is compressed, tool output is rewritten short, and the cinema's id is fictional.
On the first evening, the house voice assistant gets a web search tool: a Home Assistant script exposed to Assist, calling Tavily. The setup works on the first try, except for one checkbox. Then a movie’s showtime comes out wrong. The tool was called, but it only handed the model a one-sentence summary. Raw excerpts, with the current time placed last, give the right screening.
Two days later, same question for another film, and nothing works. Searching
by title returns film pages with no times, the model fills the gap with
plausible, wrong hours, and the search index serves a page dated two days
earlier. The extract call, meant to read the live page, returns the same
one. For dated content, an index-based search engine is the wrong tool.
The second tool queries the cinema’s API directly, which only returns upcoming screenings. Along the way, the agent stops at the guardrail that keeps it from widening its own permissions, then misses a film because of a single letter: a strict filter applied to dictated input.
The full setup is in Bob’s article and what it changes at home in Ludo’s, both in French only.