Harsh-karn/Agent-Platform-Production-Reliability ? reverse-engineered prompt

Build me a Python library for making LLM agents safe and reliable in production.

I want a set of reusable pieces that help with the messy parts of agent apps, like forcing JSON outputs into a strict schema, retrying with the exact validation error when the model gets it wrong, routing work to cheaper or stronger models based on task difficulty, and stopping execution when a budget is hit. It should also handle tool use safely, including running multiple tools in parallel, keeping the original order of results, and returning clean error messages when one tool fails instead of crashing everything.

Please also include human approval for protected actions, a basic ReAct style agent loop with a hard stop so it cannot spin forever, event processing with retries and a dead letter queue, and tracing so I can see nested calls, timing, and cost. Add a small memory system, self evaluation, and a simple multi agent debate flow too. Make it well tested and reliable.