The situation
Vandebron's AI incubator was building a multi-agent system to suggest new product ideas: market, customer and compliance research, then a summary. I led delivery.
The problem
- For weeks the agent reported that it had found nothing.
- One run cost $26 and took 15 minutes, with no output.
What I did
- Found two bugs hiding each other: a source list the search code silently read as empty, and a loop where the agent re-read its own conversation nine times per request.
- Enforced the rules in code rather than asking nicely in prompts: no number the data didn't produce, every claim linked and dated, disagreements between researchers kept visible, and legal risk routed to a person.
- Set the scope with a clear must-have versus nice-to-have list, and gated releases with 340 automated tests.
Getting an AI to find things is easy. Getting it to tell you what it doesn't know is the real job.