Field notes, experiments, and writing from the lab.
Check both meters before the job runs
Lab note: a usage gate that routes agent jobs between model providers before a cap hits. Check the weekly and the 5-hour session usage, run only if both are at or under 90%, otherwise fall to the next provider, and fall to the next provider too when the check itself errors.
Diff the live ad every night
Lab note: the design for a nightly job that snapshots what an ad platform is actually serving, compares it with what was approved, and flags any change nobody on the team made. Written as the job I'd want, with the open problems.
Five pins before an agentic measurement tool writes
Agentic measurement and ads agents (Lifesight, Snowflake AI Gateway for Advertising, Contentsquare's agent-traffic split) are arriving fast. Five pins before one gets to recommend or write a budget change: split humans from agents, own the events, check causality, human approval plus audit row, and no self-grading.
Autoloop is not a control plane
Post-deploy self-tuning for AI agents (Kore.ai Autoloop, Google's agentic Gemini, RSA Agent ID) is real progress. A loop that rewrites a live agent still needs pinned contracts, identity and authority, revoke, shadow-before-write, and audit.
DevDay control-plane checklist
Pre-DevDay operator pin: agency, economics, contracts. Score the stage against your harness before you clap.
Cost-per-task beats benchmarks when fleets call tools
Frontier week: Astra vs Gemini 3.8 Flash cost gap, Muse Spark pitching fewer tool calls. Price-per-token lies under agentic load. Measure cost-per-completed-task against a pinned schema.
The agent looped. The fighter shipped.
Clash of Steel Blades is a 2.5D Khanda vs Katana fighter in the browser. The first coding agent doom-looped on the build. Switching the worker finished it.
The agent did not get stupid. The schema moved.
When a tool schema drifts under a live agent — enums renamed, param types tightened — the agent looks broken. It is not. The contract broke. Pin versions. Fail closed.
Most of your ChatGPT traffic is hiding in Direct
GA4's AI Assistant channel is a floor. Most ChatGPT visits still land as Direct because the client strips the referrer.
Agent tabs in Telegram
Official Grok Bot still has no Telegram. I didn't install a chatbot. I made each topic an agent.
Two games, built to learn
A woodblock samurai brawler and a head-tracked rail shooter, both built as proof of concepts. The point wasn't the games. It was being a beginner again, where the agent can't fake it.
Your agent's bill isn't the model. It's the query path.
The 2026 AI cost reckoning trains your attention on the token meter. When I traced where an agent's spend actually went, the model was the easy part. The leak was the warehouse the agent queries.
The agent control plane is the actual fight
Karpathy joined Anthropic to run autoresearch at frontier scale. The same closed loop is the operator-tier curriculum now, one tier down.
Notes on how observer-agents fail
Three failure modes in agents that synthesize work you own, and why the third one breaks the whole review chain.
The shape of agents that observe their own outputs
A class of agent is emerging that doesn't do new work. It reads what you already have through a perspective you didn't take.
Standing up acidlemon, again
Why the lab notebook is back, the persona split with joelcitron.com, and how DESIGN.md became the source of truth.