Writing
Notes on MCP, agent infrastructure, open-source maintenance, and the long tail of production problems demos ignore. Some are about tools, some are about systems, and some are about what happens when the cost of writing code drops faster than the cost of understanding it.
Subscribe via RSS.
Essays
-
From 1,200 tools to 8: reshaping OpenAPI-derived MCP servers so agents can actually use them
A dump of every OpenAPI endpoint into an MCP server is technically correct and practically useless. Here's the transformation layer I've been building into FastMCP — and why picking the right 8 tools beats exposing all 1,200.
-
Pull Requests Are Dead. Long Live Context.
AI made code generation cheap without making review cheap. For open source, that shifts the scarce resource from code to context.
-
Agent Builder, Beyond The Chatbox: Introducing Augmented Infrastructure
We built an agent that does more than talk about infrastructure. It can reach into real environments, call tools through runners, and close the loop with Elasticsearch.
-
Bad Tools Make Bad Agents
Most agent failures are not model failures. They are tool-shaping failures. If your tools are generic, leaky, or poorly scoped, your prompt turns into a compensating control system.
-
Why We Shipped Performance Presets Instead of Better Documentation
Elastic Agent had accumulated a pile of tuning knobs that only experts could use well. Performance Presets were a product decision to package judgment, not just publish more YAML advice.
At Pydantic
Product notes, launch essays, and practical field guides written for Pydantic.
-
The best Langfuse alternatives in 2026, honestly compared
-
The 7 best LLM evaluation tools in 2026
-
The best OpenTelemetry backends in 2026, honestly compared
-
The best prompt management tools in 2026
-
The best Sentry alternatives in 2026, honestly compared
-
Do evals the Airbnb way
-
Fork the loop
-
Focus on evals with Logfire
-
Score freely
-
The best AI platform for building agents on Kubernetes in 2026
-
The best AI agent optimization platforms in 2026
-
Ten agents, ten clouds, one answer
-
The agent outgrew your laptop
-
The agent writes faster than you can review
-
A research agent, three ways
-
Your traces already know how to fix your prompt
-
Score Freely with Pydantic Logfire
-
One key in, no keys out
-
The average run is lying to you
-
You perfected the wrong agent
-
When you do not know what you are looking for
-
The OOM that was not your agent
-
The pod that did not survive Tuesday's deploy
-
When the slow thing is not the model
-
Agents are the new services