Pydantic

Our Head of Product, Bill Easton sits down with Kayvon Beykpour and talks about the secret sauce (it's Macroscope) behind how we've delivered against our Pydantic Logfire roadmap 5x faster in 2026 than we did in 2025 with the same team. Check it out 👇

Pydantic AI agents now run on Jev, the classifier from TypeSafe AI. Jev doesn't write text. It answers typed questions. So the `output_type` you already wrote is the question: each field is something Jev answers, and the answer comes back as your model, with a confidence per field. Attach tools and Jev picks which one the text calls for. It calls the ones that take no arguments itself, and raises the ones that need arguments so `FallbackModel` hands that step to a language model. On 120 support tickets, 115 never left Jev. The classifier does the part it is good at, and the language model picks up the part that needs writing. Measured on those same 120 tickets with Pydantic Evals: 6x faster than gpt-5.6-luna, at about a third of the cost. Accuracy stays out of it, because at that sample size the models don't separate. Speed and cost are the gaps that hold up. Link for the docs in the comment section.

Starting in 40 minutes: Marcelo Trylesinski on what that feedback loop actually needs at AGNTCon + MCPCon Europe in Amsterdam. Self-healing agents sound inevitable: an agent fails, finds the cause, proposes a fix, validates it. Without observability all of it is just guessing. @Marcelotryle on what that feedback loop actually needs. 10:20 CEST, G104 + G105. #AGNTCon #MCPCon

Our Security & Compliance Lead Matt Franz spoke at DC State of the Stack (formerly DevOpsDays DC) in Washington. Go-ing Fast with MCP: Rapid Development of MCP Servers for Security Agents. The argument: MCP is how security agents reach your tools, but a Python MCP server drags in a dependency tree and a slow start every time an agent spawns it. Even with uv. Go gives you a single binary, fast startup, and typing that maps cleanly onto tool schemas. Matt covered wrapping SIEM, SOAR and EDR APIs into tool definitions an LLM can actually use, bounding what goes in and what comes back, and testing servers end to end against real models. Missed the talk? Slides and the architecture guide are here: https://lnkd.in/deqfZk4Y

Self-healing agents sound inevitable. An agent fails, works out what went wrong, proposes a fix, validates it, improves the system. Without observability, that's just guessing. Marcelo Trylesinski is at AGNTCon + MCPCon Europe in Amsterdam speaking about this: what a real improvement loop needs underneath it, built on traces, evaluations, approvals, and rollback. He'll follow a single agent run through prompts, model calls, tool selection, MCP servers, validation, policy checks and downstream APIs, then show what evidence you actually need to reconstruct what happened. Self-Healing Agents Need Observability Friday 18 September, 10:20 CEST, G104 + G105 (Level 1) If you're there, grab him to talk MCP, agents, and what observability looks like in production. #AGNTCon #MCPCon

If you trace LLM calls in one tool and the rest of your app in another, agent runs that don't throw errors are hard to debug. You end up correlating timestamps across tabs to find the query that fed the model bad context, and you pay for and manage multiple observability stacks. On Sep 30, Jameson L. will show you how Pydantic Logfire records the model call and the rest of your request in one OpenTelemetry trace, and where evals fit once you have those traces. Save your spot: https://pydantic.io/UM58j

Picking an image model takes trying a few. That used to cost a new SDK, a new request shape, and a new response to unpack before you could compare outputs. The new 𝗜𝗺𝗮𝗴𝗲𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗼𝗿 in Pydantic AI gives one interface across OpenAI GPT Image, Google Gemini, and xAI Grok Imagine, so trying a different model is a change of string. Settings like dimensions and aspect_ratio are typed, and you get warned about ones a provider will ignore or override, rather than finding out from the image that comes back. Turn on instrumentation and every generation lands in Pydantic Logfire beside your agent and tool spans - prompt, image, and latency in one trace. New from David S. on Pydantic AI: https://pydantic.io/mLFGf

Full Monty

Pydantic AI Harness now ships Amazon Web Services (AWS) LambdaDurability. It checkpoints every model request, tool call and MCP call as an AWS Lambda durable step, so a retried invocation resumes from the last completed step instead of starting over. One capability, two decorators, and an agent execution that can span up to a year without compute charges for on-demand functions. An agent that resumes is a different thing to operate than an agent that restarts.

Last call to register and join us live for this online session!