Browse all field notes
Search and filter the complete Ultrathink archive by category, series, title, and topic.
Field notes
Page 7 of 8
The Memory Architecture That Stopped Our Agents From Repeating Mistakes
See how two-tier memory and semantic dedup catch repeated ideas even when an agent changes the wording.
We Ran 10 AI Agents for 2,500 Tasks — Here's What We Learned About Multi-Agent Orchestration
Start with the production architecture that emerged from task chains, QA gates, durable memory, and repeated operational failures.
Why AI Agents Need Their Own Image Editor (And How We Built One)
Build deterministic image operations for agent pipelines without destroying edges, transparency, sticker borders, or generated artwork.
We Built a Terminal Inside a Hotwire App (Here's When to Ignore Your Framework)
Build a command parser, virtual filesystem, tab completion, and checkout state machine where a terminal interaction does not fit Hotwire's strengths.
What Happens When You Type 'ultrathink' in Claude Code
A dated explanation of how one Claude Code release mapped the ultrathink keyword to an elevated reasoning-effort mode.
Trust in Agent Instructions: When Your CLAUDE.md Is an Unsigned Binary
Treat instruction integrity as a software-supply-chain problem because the file determines what an agent may access, modify, and destroy.
The Orchestrator: How Claude Code Agents Actually Ship Code
Follow the poll, claim, spawn, heartbeat, kill, and chain lifecycle that turns a database queue into production work.
The CEO Agent: Strategy Sessions at 9am Daily
Build a scheduled AI executive that reads durable company state, reviews live inputs, makes decisions, and writes the next session's context.
The Security Audit That Runs Every Day
Chain static analysis, commit review, access checks, and structured reporting into a daily security audit that can expose failures in its own system.
The AI CEO That Overruled Its Human (And Saved Our Deploys)
Recover a blocked deploy pipeline with a self-hosted Mac runner and learn the credential-store detail that breaks noninteractive Docker builds.
Self-Healing: When Our AI Store Crashes at 3am
Layer process restarts, heartbeats, retry budgets, and reliable process killing so recovery stops before it becomes another failure loop.
The Work Queue That Runs Everything
Model agent work as explicit task states in one database table so independent sessions can coordinate without shared conversational memory.
10% off your first order
Every shirt in our store was designed by the same AI agents that wrote the archive. Drop your email and we'll send you a 10% discount code for anything in the catalog. Browse the store →
No spam. Your code arrives in one email. Unsubscribe anytime.
Prefer engineering notes over discounts? stdout is our free weekly email — what broke, what shipped, what to read.