IDEAS, DECISIONS & LESSONS

Journal

Notes on AI, architecture and building useful products.

279 articles

20 Sep 2026 · 5 min read

agents

Visa, Mastercard, and Google Just Built Three Different Ways to Leash an AI Agent's Wallet

AP2, Trusted Agent Protocol, and Agent Pay all shipped inside thirteen months to let AI agents pay for things safely. A September 2026 paper found ways to make an agent sign for the wrong amount anyway — with success rates from 1.2% to 100% depending on the model.

20 Sep 2026 · 4 min read

Your AI Coding Agent Just Ran a Branch It Swore Was a Commit

Claude Code, Codex, Copilot, and Gemini CLI all pin plugins to a commit SHA for safety. A Git ref-resolution quirk means that promise was never actually true — and the four vendors responded four different ways.

19 Sep 2026 · 5 min read

agents

OpenAI's Agents API Sandbox Pricing Has Three Blind Spots — Here's How I'd Budget Around Them

OpenAI opened its managed Agents API on September 10. The sandbox pricing looks simple until you try to actually forecast a bill. A build-vs-buy breakdown for tech leads.

19 Sep 2026 · 5 min read

Your Red Team's Fake Company Can't Share a Real One's Name

Gemini breached three real companies during a sanctioned red-team exercise because the test target had the same name as a live business. The lesson isn't about rogue AI — it's about namespace hygiene.

19 Sep 2026 · 4 min read

agents

The Reel Where AI Transfers Your Shares Is a Joke. The Two Real Incidents Behind It Aren't.

A viral reel jokes that asking AI to run your marketing team ends with it transferring your shares. Two real 2025-2026 incidents — a deleted production database, a $250K-plus token transfer — show what actually goes wrong, and why the fix isn't a smarter model.

19 Sep 2026 · 4 min read

DoorDash Put a Price on Multi-Agent Code Cleanup: $4.79 a Flag

DoorDash ran 50 stale feature flags through a Sonnet-plus-Opus agent pipeline and published the numbers. The interesting part isn't the cost — it's what made the 90% success rate possible.

19 Sep 2026 · 5 min read

agents

Fable 5.1's Cache Reads Dropped to $0.25 — But My Agent Loop Still Got More Expensive

Anthropic cut Claude Fable 5.1 cache reads 75%. Artificial Analysis found max-effort runs cost 20% more anyway, because output tokens jumped 1.7x. I ran the math on our own agent loop to find out which story was true for us.

18 Sep 2026 · 7 min read

agents

Tsinghua Turned 'Build Your Own LLM' Into Homework. Here's How I'd Actually Do It

On September 18, Tang Jie — Tsinghua professor and Zhipu AI chief scientist — opened a course where students build a tokenizer, Transformer, Triton kernel, and a self-evaluating agent from scratch. I checked the sources and wrote the practical version for anyone without a Tsinghua lab budget.

17 Sep 2026 · 4 min read

Google's Gemini 3.8 Flash Cyber Beats Every Frontier Model at Finding Bugs. Here's What That Changes in a CI Pipeline

Gemini 3.8 Flash Cyber hits 86.2% on CyberGym, beating Mythos 5 and GPT-5.6 Sol, and Chrome's security team says it produces 2.6x more correct patches than larger commercial models. I looked at what a purpose-built security model actually changes for teams wiring AI into AppSec workflows.