Google Research introduced Agile Deliberation, a human-in-the-loop framework for building classifiers around subjective visual concepts such as "healthy food" or "paid-to-play" clickbait, released as open-source code under Apache 2.0. The system operates in two stages within a Jupyter/Colab notebook, leveraging Gemini 2.5 Pro and Flash via API without fine-tuning, alongside CLIP embeddings indexed by FAISS (default 260 MB demo index, scaling to 13 GB for full LAION-400M). Stage 1 decomposes a vague concept into up to three unit concepts with candidate positive and negative subconcepts, presented with retrieved images for user validation. Stage 2 replaces uncertainty sampling with semantic borderline discovery: Gemini 2.5 Flash generates ambiguity queries (e.g., "salads with heavy mayo dressings"), retrieving 50–100 images that are deduplicated, clustered via dictionary learning, and prioritized by a UCB multi-armed bandit favoring clusters with user corrections. Users label at least five images per shared ambiguity; an LLM synthesizes feedback into candidate definitions, with a greedy update retaining the highest-F1 prompt. Evaluated across 18 live sessions (9 participants, ~90 minutes each) on a proprietary 100M-image dataset and WebLI, Agile Deliberation achieved 0.59 F1 on "paid to play" (+11 points over zero-shot Flash's 0.48, precision +22 points to 0.57, recall -17 to 0.63) and 0.58 F1 on "healthy food" (+10 points over 0.48, recall -15 points). It outperformed Modeling Collaborator (CVPR 2024) by 6–8 F1 points while requiring only ~45 minutes of human effort versus ~100 labels plus chat. Users reported significantly lower effort and stress compared to manual prompt engineering.