Archive

Wooden Pinocchio puppet collapsed on its back, limbs splayed: the eval that fell over when it stood on fabricated history.

ai-agents· llm· observability· blueclaw

How to Evaluate LLM Agents: 5 Patterns That Catch Failures

Eval shortcuts fail on honesty-trained models. Five assertion patterns that survived production.

July 11, 2026 · 12 min

Empty station hall at dawn, tracks running out toward the horizon: by the time the train leaves, every switch that matters has already been set.

rag· ai-agents· llm

RAG for AI Agents: 6 Decisions That Make or Break Retrieval

Six retrieval decisions for agent RAG, each settled by a benchmark, not a vibe. Defaults included.

July 10, 2026 · 10 min

A search result returning a single focused paragraph instead of a noisy text window.

mcp· ai-agents· python· llm

How One Search Change Eliminated an Entire Agent Step

pdf_search went from a pivot tool to a terminal tool. The thesis still holds.

July 04, 2026 · 8 min

Two people reviewing handwritten notes between two laptops: the collaborative review loop between developer and LLM evaluator.

mcp· ai-agents· production-systems· python

How to Evaluate an MCP Server With an LLM: 17 Bugs Found and Fixed

Your tools work. That doesn't mean an agent can use them. Eight rounds. Seventeen bugs.

June 27, 2026 · 12 min

A reader's finger resting on a single line of an open book, standing for an agent locating the one passage it needs instead of reading every page.

ai-agents· mcp· rag· llm

How AI Agents Should Read PDFs: 5 Patterns That Survived Production

How agents should navigate documents at production scale, from 32,000+ downloads of one MCP server.

June 24, 2026 · 11 min

aws· serverless· python

How I Built a Serverless Newsletter on AWS for Under $1/Month

Seven Lambdas, two SQS queues, one DynamoDB table. SES for sending. No per-subscriber fee.

June 20, 2026 · 10 min

A 1940s switchboard operator routing tangled cables on a wooden patch panel: too many wires to plug, too few hands to plug them.

mcp· python· ai-agents· production-systems

MCP Tool Sprawl: How I Cut 69 Tools to 43 With a Decorator

Why more MCP tools make agents worse, and the pattern that fixes the surface.

June 13, 2026 · 8 min

A rocket lifting off the pad in a billow of golden-lit exhaust: a payload built on the ground, now going live and heading somewhere remote.

mcp· security· ai-agents· llm

How to Deploy a Python MCP Server: Remote HTTP, Auth, and Docker

Take the notes server from localhost STDIO to a secured, containerized HTTP service you can run on any host.

June 09, 2026 · 11 min

An antique Underwood Standard typewriter on display: a simple mechanical tool that still does its one job exactly right.

mcp· ai-agents· rag· llm

Section-Level RAG: Why BM25 Beat Hybrid Search in My Benchmark

Hybrid wins at page grain. BM25 wins at section grain. Granularity decides.

June 06, 2026 · 10 min

A hand flipping through an open library card catalog drawer: a curated index that takes you straight to the right card instead of every card.

mcp· ai-agents· rag· python

Section Chunking vs Page Chunking for AI Agents: ~6 Fewer Tool Calls Per PDF Query

Page-mode PDF search costs 2 to 6 extra tool calls per query depending on the document. Section-aware search delivers...

May 30, 2026 · 10 min

Wooden Scrabble tiles on a table spelling the word FREE: something useful that costs nothing extra.

mcp· ai-agents· python

Your LLM Is Free QA for Your MCP Server

Every Claude Desktop session using your MCP server is a free QA pass. You just have to listen to what the LLM is tryi...

May 23, 2026 · 10 min

A fisherman casting a wide net across calm sunset-lit water: catching everything in one sweep instead of one fish at a time.

ai-agents· production-systems· llm· blueclaw

I Built CI for My AI Agent (It Catches What You Miss)

A prompt change broke my agent silently. Behavioral CI caught 4 regressions before they shipped.

May 16, 2026 · 8 min

Latest — recent dispatches