Pydantic

"What I want most: software that helps me express, formalise, and include my opinions rather than enforcing its opinions on me." In the 3rd and final part of Lies, Damn lies, and the Box Model series, 🧐 Laura Summers talks about the design process, intuition, cargo cult prompting, and the promises that will survive the froth. https://lnkd.in/dWNz4i5j

If your AI observability vendor is making you sample traces to manage costs, that's not observability. That's a blind spot you're paying for. We compared AI observability pricing at production scale. The cheapest competitor costs 8x more than Logfire. The most expensive, 100x. If you're building production AI applications and paying more than $2 per million spans, run the numbers. Link in the comments.

Improving your agents in production without redeploying is an actual problem. Catch Samuel Colvin workshop at AI Engineer on April 8th (10:40am UTC+1) showing how Managed Variables + GEPA turn your Logfire traces into a continuous improvement loop for your Pydantic AI agents. No redeploy. No downtime. Eventually, no human required. And if you want to continue the party, join us at 6pm at the @atomico office in Rathbone St for the PyAI London meetup with Samuel Colvin, Shifra Williams, David Hewitt (PyO3), Marlene Mhangami , and Pablo Galindo Salgado. Hosted by Laís Carvalho .

An AI agent named a container MS13. Within 30 minutes, the CISO of a major hyperscaler was paged. A data center cage was locked out. The task was: spin up some Docker containers. A model that discovered a contradiction between the Bible and biology textbooks. It did not move on. It consumed an entire data center's worth of compute debating the Immaculate Conception with itself. Every source validation heuristic it used was correct. The outcome was not. This talk is full of stories like this. Real incidents from production AI systems, research training runs, and early computer vision, told by someone who was in the room when they happened. The point isn't that AI breaks in wild ways. It's that you couldn't debug any of it without observability. Structured traces, end-to-end logs, human-readable outputs. That's what turned "something is very wrong" into "here's exactly what happened and why." Full talk on YouTube. Link in comments.

Nobody opens an observability tool and thinks "wow, what a sidebar." But bad navigation creates friction on every interaction, and that friction compounds. On Part 2 of Lies, Damn Lies, and the Box Model, 🧐 Laura Summers wrote about how was to redesign Logfire's navigation bar. The trade-offs, and what "directionally correct" means in practice. Link in comments. Part 3 (and final) out tomorrow.

Pydantic AI agents now have native support for durable execution via Restate. Automatic retries, persistent sessions, human-in-the-loop approvals, multi-agent orchestration that survives failures. Production-ready agents without rebuilding your stack.

Starting a Python interpreter takes ~1 microsecond with Monty. A sandbox call takes ~1 second. That's 6 orders of magnitude. And no external service, no cgroups, no VM. Samuel Colvin explains why he built a Python interpreter in Rust from scratch for AI agents on his PyAI talk. Link in comments.

Our lead designer 🧐 Laura Summers paired with Claude Code to debug a CSS layout bug: a scrollbar that only appeared on a specific page in Logfire. Claude could grep the codebase, trace component hierarchies, and suggest fixes faster than she could type. But it couldn't see what was wrong with the page. The root cause took her an hour to find. She wrote about what it taught her about AI pair programming, visual debugging, and why 'seeing' is still a human-only skill. Three takeaways: 🤖 AI is great at searching code. It's weak at seeing design. 👾 CSS has no 'easy' problems, just problems you haven't hit yet. 🔍 Always push past the first fix to find the real cause. Link for the full write-up of part one in the comments. Part 2 out tomorrow.

If your AI observability vendor is making you sample traces to manage costs, that's not observability. That's a blind spot you're paying for. We compared AI observability pricing at production scale. The cheapest competitor costs 8x more than Logfire. The most expensive, 100x. If you're building production AI applications and paying more than $2 per million spans, run the numbers. Link in the comments.