AgentMon: Building Step-Level Tracing + Expert Review for PydanticAI Agents

AgentMon is a lightweight framework I built to make agent behavior inspectable after a run. I’ll demo how a PydanticAI “deep research” agent is instrumented with a simple decorator to capture a structured step trace (tool calls, LLM calls, state snapshots, final output) into Postgres. Then I’ll walk through the post-run augmentation pipeline that maps raw steps into an agent-specific ontology and summarizes long outputs so runs are easier to review. Finally, I’ll show the web UI timeline where reviewers can score/label individual steps and whole runs, and how those annotations feed basic analytics across runs. Early stage but iterating fast.