Typed agents, fully traced.
PydanticAI brings type-safe agents and structured outputs to Python; Obsivara tells you what each run costs and where it breaks. Every agent run traced through tool calls, model requests, and output validation — each carrying tokens, latency, and dollars — so a run that retries on a validation failure points to the exact step. PydanticAI is OpenTelemetry-native; point its instrumentation at Obsivara, out of the request path.
PydanticAI, fully observable.
Run and tool tracing
Every agent run captured as a span tree — model requests, tool calls, and their inputs and outputs, with timing at each step.
Structured-output visibility
Output validation and retries traced, so you see when a model returned data that failed the schema and how many attempts the run took to recover.
Per-agent cost attribution
Tokens and dollars rolled up per agent, run, and model, so you know which agents and which validation-retry patterns drive spend.
Failure attribution and health
Validation failures, tool errors, and model failures classified and traced to the step that broke, with health scoring and predictive alerts.
Three steps. No code changes.
Point its OpenTelemetry at Obsivara
PydanticAI is OpenTelemetry-native (via its Logfire/OTel instrumentation) — configure its exporter to Obsivara's OTLP endpoint, or wrap your app with the Obsivara SDK. Your agents keep calling models directly; Obsivara records each run out-of-band, with no proxy in the path and no added latency.
Run your agents
Traces flow as runs execute tools and validate outputs — no changes to your agent or model definitions.
See it in Obsivara
Agents appear in the inventory and knowledge map with per-run cost, latency, and health within minutes.
Signals tracked out of the box.
- →Agent run and tool traces
- →Structured-output validation and retries
- →Tokens and cost per agent and run
- →Latency per step and model
- →Error rates and failing steps
- →Health score and predictive alerts
Common questions.
How does Obsivara connect to PydanticAI?
PydanticAI emits OpenTelemetry (the same instrumentation that powers Logfire); point its OTLP exporter at Obsivara, or use the Obsivara SDK. No hand-instrumentation of your agent code is required.
Can Obsivara show structured-output retries?
Yes. Output validation and the retries it triggers are traced, so you can see which agents waste calls recovering from schema failures.
Do I need to change my PydanticAI agents?
No. You configure the OTel exporter or the Obsivara SDK; your agents, tools, and output models stay exactly as they are.