Observability and execution traces
When an AI workflow fails, I want to follow the steps that led there. The prompt, context, tool calls, and timing can all matter.
A record you can use
A trace should let me reconstruct a run and compare it with another. That is the problem behind my trajectory-service work at USLBM.
Trusting the telemetry
Missing events, changing schemas, and failed messages can leave gaps in that record. Ludex explores the health of the telemetry pipeline itself.
Open questions
- Can one run be reconstructed from the instruction through the tool calls?
- What is missing when the signal arrives incomplete?
- How do you tell a model failure from a pipeline failure?