Parallel large language model (LLM) agents explore multiple solution trajectories to tackle complex tasks, yet valuable discoveries often stay isolated within individual execution histories, making them inaccessible to other agents. Existing sharing mechanisms usually select or summarize information before the actual need becomes clear, risking the loss of details that become crucial later.\n\nTo address this, we introduce Ask on Demand (AOD), a communication framework that lets agents request targeted peer assistance exactly when a specific information need arises. Once a request is approved, AOD retrieves relevant evidence from peers' execution histories, chooses an appropriate peer, and generates a response tailored to the request.\n\nWe train a lightweight gate to decide when a request is worthwhile. The gate is learned from paired team continuations, with supervision derived from the final team outcomes and the inference cost incurred.\n\nExperiments on the AppWorld and WebShop benchmarks show that AOD consistently outperforms independent parallel execution in task success rate, while achieving lower inference cost than existing information‑sharing baselines.\n\nThese findings highlight that aligning peer communication with the evolving information needs of agents leads to more effective and efficient collaboration.\n\nReview: AOD’s demand‑driven query mechanism mitigates the noise introduced by premature sharing, and its lightweight gating balances cost and benefit, offering a practical blueprint for collaborative parallel agents.