When an agent gives a wrong answer, the cause is usually several steps back: a tool returned something unexpected, the model misread it, and every later step built on the mistake. Observability means recording each model call, tool call, input and output as a connected trace, so you can follow a run step by step instead of guessing from the final answer.
These guides start with tracing an agent's tool calls with OpenTelemetry-compatible spans. The next one uses those traces to diagnose a common failure, an agent that repeats the same step without making progress, and shows how to stop it. The last covers the server side: metrics, logs and health checks for the MCP servers your agents depend on.
The code in these guides runs on Promptise Foundry, an open-source Python framework for AI agents and MCP servers.