Field notes from the studio
Practical write-ups on shipping AI agents, RAG systems, and data infrastructure into production. No fluff, just what worked.
Your Agent Traces Are Lying to You: The Race to Standardize Agent Observability
OpenTelemetry just shipped semantic conventions for agent spans, Datadog adopted them natively, and the Linux Foundation stood up a new governance body for agent infrastructure. Here's what's actually usable today, what to skip, and why your traces still might not tell you why the agent did what it did.
Read articleAll articles
15 posts
Your Eval Gate Is a Rubber Stamp: Why LLM-as-Judge Misses Most Long-Horizon Agent Failures
A growing body of production and academic evidence shows that LLM judges validated on single-turn agreement scores are structurally blind to the failures that actually break long-running agents. Here's what the data says and what to build instead.
Read more
Why Multi-Agent Systems Fail: The First Real Taxonomy of Production Breakdowns
A UC Berkeley study of 1,600+ multi-agent execution traces finally gives the industry a rigorous vocabulary for why agent fleets break in production, and it maps directly onto fixes teams can ship this quarter.
Read more
The Sandbox Is the Product: Why Agent Execution Isolation Just Became Its Own Layer
Agents now write and run their own code. That means the box they run it in matters more than the prompt that got them there. Here's what changed, what broke, and how the isolation decision tree actually works.
Read more
Agent Memory Is Having Its RAG Moment, and Filesystems Are Winning
A wave of new benchmarks, SDKs, and native API primitives has turned agent memory into its own architectural layer. The early data says the simplest approach, plain files, is beating purpose-built memory databases.
Read more
Stop Giving Agents a Service Account and Calling It Done
Agent identity is turning into its own discipline, separate from prompt safety and tool architecture. The default pattern of borrowing a human's full-scope token is now a documented anti-pattern, and the fix looks a lot like OAuth delegation, not a policy memo.
Read more
Code Mode: When Your Agent Stops Calling Tools and Starts Writing Programs
A quiet architecture shift is moving agents from exchanging JSON tool-calls with the model to writing code that calls tools directly. The token savings are real. So is the new infrastructure bill: sandboxing, artifact storage, and script reuse that most teams haven't budgeted for.
Read more