Show HN: I ran every Claude agent turn through the Batch API (eran.sandler.co.il)

A developer has created a "batching-harness" that allows each turn of an AI agent to process through Anthropic’s Batch API, which offers a significant cost-saving of 50%. This experimental setup, built in Python, combines a minimal sandbox environment with a user-friendly terminal UI, making it straightforward to evaluate how the Batch API's asynchronous processing impacts user experience. However, the results revealed a major downside: processing times can extend to two minutes per turn, which greatly reduces the efficiency of interactive AI agents, transforming a simple task into a lengthy endeavor. The findings suggest that while batching can be cost-effective for specific use cases—like running multiple agents concurrently or during non-time-sensitive tasks—it may not be beneficial for individual users requiring quick responses. Interestingly, it was noted that certain models (like Haiku) performed worse in batch scenarios compared to slower models due to their faster execution times. This insight challenges conventional wisdom about using cheaper models for offline tasks, advocating instead for a strategy that involves batching multiple requests from diverse agents to optimize both cost and performance. The developer hopes to further explore this concept by implementing a local proxy capable of smartly managing batching across various AI agents in the future.