Observability

Every assistant turn produces a trace tree. Use observability to understand why an agent said something, which tools ran, what memory was retrieved, and where latency or errors occurred.

Span types

Hyponema records spans for:

• Tasks.
• LLM calls.
• Tools.
• Functions.
• Retrieval.
• Memory operations.
• Guardrails.
• Scores and evaluations.
• Errors.

What to inspect

For a session, review:

• The transcript.
• The LLM pass hierarchy.
• Tool inputs and outputs.
• Memory retrieval context.
• No-go-zone guardrail results.
• Latency and provider metrics.
• Error payloads and retry behavior.

Production workflow

Use traces when:

• A customer reports a bad answer.
• A tool appears not to run.
• A persona change changes behavior.
• Latency rises.
• A guardrail blocks or redirects unexpectedly.
• You are validating a new provider stack.

Observability is also the fastest way to confirm that memory is being retrieved, but not over-injected.