Blog
How I Improved an AI Agent from 40% to 60% — With A/B Test Data
Same model, same test cases, 20% better results. 7 out of 8 fixes were pure code, zero LLM cost. Here's exactly what I changed and why it worked.
[ AI_OPERATOR ]
AI Operator track — production AI agent engineering. Validation loops, tool-call reliability, completion ownership, context engineering, cost forensics, and observability for systems that ship and stay shipped.
Long-form writing and video on production AI agent engineering — the 90% of work that happens after the model gives you a demo that works.
Topics: validation loops, tool-call completion ownership, context engineering, retry budgets, cost forensics, and observability for systems that ship and stay shipped.
Companion to the Harrison AI Operator YouTube channel. Blog and video are listed below — most recent first.
Blog
Don't Pick One AI. Run Three Against Each Other.
Don't bind to a single AI. Run three in competition, make the final call yourself, and let results judge everyone. The operating model for staying valuable in the AI age.
Video
Why Your AI Agent Keeps Failing — The 90% Problem
Your AI agent probably isn't failing because the model is weak. It's failing because you're fixing the wrong layer. The other 90% — context, memory, validation — is where production breaks.
Video
The 90% AI Agent Problem (Podcast Episode)
Building an AI agent that works is easy. Building one that keeps working is where most teams fail. Long-form breakdown of the four-layer failure model.
Blog
The 90% Problem: Why Most AI Agents Are Still Broken
Building an AI agent that works is easy. Building one that doesn't break is 90% of the work. Here's what that 90% actually looks like — from leaked source code and production A/B data.
Blog
Claude Code + Codex Plugin: Two AI Brains, One Terminal
How to use OpenAI's Codex plugin inside Claude Code — turning Claude Opus and GPT-5.4 into a dual-brain coding system. Setup, commands, rescue workflows, and when each brain wins.
Blog
Claude Code Deep Dive Part 4: Why It Uses Markdown Files Instead of Vector DBs
Claude Code's memory system looks simple on purpose. This piece breaks down the tradeoffs behind Markdown memories, Sonnet side-queries, and the decision to avoid vector databases.
Blog
Claude Code Deep Dive Part 3: The 5-Level Compression Pipeline Behind 1M Tokens
Claude Code doesn't just stuff conversations into a 1M-token window. It uses a 5-level compression pipeline, cache-aware edits, and a final autocompact fallback to keep sessions alive.
Blog
Claude Code Deep Dive Part 2: The 1,421-Line While Loop That Runs Everything
Inside query.ts — the 1,729-line async generator that is Claude Code's beating heart. 10 steps per iteration, 9 continue points, 4-stage compression, and streaming tool execution. With line numbers.
Blog
Observability and Billing for AI API Calls: A T-Shaped Architecture
AI API calls are unlike ordinary RPC: per-request cost varies 100×, tokens and models are first-class, streaming muddies timing, caching changes the pricing. A T-shaped instrumentation architecture — shared stem, specialized arms — that handles tracing, billing, and cost analytics without any of them contaminating the others.