Skip to main content
Kindred is a reliability and debugging platform for AI agents. It helps teams understand production agent behavior, discover recurring patterns, define expected behavior, and reproduce failures. Kindred is built for engineers and teams building multi-step, tool-using, or multi-agent systems.

The problem

AI agents can behave differently across runs. They may choose the wrong tools, take unexpected paths, skip important checks, or produce inconsistent outputs. Normal logs show what happened in one run, but they often make it difficult to see whether the same issue is recurring, which behaviors are common across production traffic, or where behavior first changed between two runs.

How Kindred works

Kindred connects to the traces your agents already produce, then organizes them around agent behavior:
  1. Connect existing traces through Langfuse.
  2. Explore real agent sessions and logical turns in Behavior Search.
  3. Use Clustering to discover recurring behaviors and failure patterns.
  4. Use Expectations to monitor whether agents follow important behavioral requirements.
  5. Use Replay to rerun a production turn and identify the first meaningful divergence.

Core concepts

  • Session: a conversation or workflow containing one or more turns.
  • Turn: one logical agent invocation.
  • Step: an LLM call, tool call, tool result, or other operation.
  • Expectation: a behavioral requirement.
  • Cluster: a group of similar agent turns.
  • Divergence: the first meaningful difference between an original and replayed run.

Where to start