Pydantic webinar: trace coding agents in Logfire to compare cost and quality
samuelcolvin · x · 2026-10-29
Pydantic announced a free Oct 29 webinar, "Closing the loop: using coding agent traces in Pydantic Logfire," led by AI FDE Duncan McKinnon.
The premise: coding agents do real work daily but leave almost no record beyond commits and PRs, so when results are bad you lack the data to diagnose or compare tools.
The workshop covers:
- Tracing any coding harness into Logfire — full sessions, turns, model calls, tool runs, subagents
- Querying the OpenTelemetry data via the Logfire MCP server
- Building metrics to compare harnesses on cost, output and quality — e.g. which harness is best at code review, whether a cheaper model matches, whether agents waste millions of tokens or re-explore the same tasks
- Turning traces into better skills, evals and self-improvement loops
More from coding & agent
- Qwen3.8-27B spends 21 days and 180 subagents failing to beat llama.cpp — udmrzn · 2026-10-29
- Runway Dev launches on Codex for autonomous API integration — runwayml · 2026-10-29
- LangChain: evidence graphs boost agent accuracy 27% and cut cost 58% — LangChain · 2026-10-29
- ClickUp Ships Agent Browser Control That Fills Forms and Orders Pizza Live — mathemagic1an · 2026-10-29
- Why run agents on a Mac mini instead of a 70% cheaper Linux box? — willdepue · 2026-10-29
- AI/ML API launches Jev Router, an MCP connector that routes tasks by difficulty to cut costs — testingcatalog · 2026-10-29