Has anyone tried Freeplay for analyzing multi-agent conversations?

I've been evaluating Freeplay as a potential solution for a specific, growing pain point in our LLM operations pipeline: the systematic evaluation and debugging of multi-agent conversation flows. We're moving beyond simple prompt chains into more complex, stateful agentic workflows (think customer support escalations, coding assistants with specialized sub-agents, etc.), and our current tooling is failing us.

The core issue is that analyzing a single LLM call's inputs/outputs is trivial compared to dissecting a full conversation between multiple agents, each with their own system prompts, tools, and context windows. When a multi-step workflow produces a poor final outcome, root-cause analysis is a nightmare. Was it the orchestration logic? Did a specific agent misinterpret the state? Did a tool call fail silently? We need to trace the entire conversation graph, not just a linear log.

My initial foray into Freeplay's documentation suggests it has features for this—specifically the "Session" concept and the ability to tag/track custom metadata across a chain of calls. However, the marketing materials are predictably vague on the gritty details of how this works at scale for complex, non-linear interactions.

I'm looking for concrete, operational feedback from teams who have deployed it in production for this specific use case. My primary questions are:

* **Trace Visualization:** How effectively does Freeplay reconstruct and display the actual graph of a multi-agent conversation? Can you visually follow the path of a session ID across different agents, models, and tool calls? A screenshot or description of the UI would be more valuable than a feature list.
* **Evaluation at Scale:** We run thousands of these conversations daily. Can you define and run automated evaluations (quality, cost, latency) on the *entire session* as a unit, not just individual LLM calls? For example, scoring the final user satisfaction while also segmenting cost by agent.
* **Integration Overhead:** What was required to instrument your agent framework? If you're using LangChain, LlamaIndex, or a custom orchestrator, did you have to wrap every agent call, or was there a cleaner middleware approach? Example code snippets for instrumentation would be extremely telling.
* **Cost Attribution:** This is a major FinOps concern. Does Freeplay allow you to tag costs (e.g., token usage, provider cost) to specific agents or branches within a session? We need to answer questions like "How much more expensive is our 'Senior Engineer' agent compared to our 'Triage' agent?"

I am not interested in general "it's great for observability" platitudes. I need to understand the practical limits, the friction points, and whether the tool actually delivers actionable insights for improving the reliability and cost-efficiency of complex agent systems. If you've hit limitations—such as poor support for asynchronous agents, max session length issues, or inadequate granularity in the trace data—that is equally, if not more, important to share.