Follow on TelegramLearning is a stochastic process. I do the trial and error, you read what worked.
Security
A flaw in the Brotli library, inside Node.js and Bun, can crash AI agents and might let attackers run code. Since October 27, CISA has listed it as exploited. You need Node 22.23.4, 24.22.1, or 26.12.1 (or later on the same release line), Bun 1.4.4, and Claude Code 2.1.308.
Explainers
You know what you load on every request, you turn off what you don’t use, and you move hard rules into permissions: the monthly routine, with measurements from my setup and a prompt that does the inventory for you.
Newsletter
On Monday morning, you’ll know:
Tutorial · Tools
Two keys in settings.json turn off auto-compaction and idle compaction: you run compaction yourself when you need to.
Explainers
In a workflow, you write the steps in code. In an agent, the model generates them one turn at a time. I explain what the difference costs (about 4 times the tokens of a chat, according to Anthropic) and the rule for choosing before you start.
Explainers
After compaction, an agent broke an explicit constraint in 30 runs out of 100, compared with 0 before (Wang et al., July 2026): the summary keeps the task and loses the rails.
Explainers
The frontend has relied on three monopolies (interface, persuasion, and decision), and it’s losing all three. With personal agents, value moves to APIs, MCP, A2A, x402, and ERC-8004, and a website has to become queryable by a machine, not just convincing to a human.
Explainers
When AI agents search, compare, and pay on our behalf, the web moves from user experience to machine experience: what happens to attention, work, and social control, from x402’s pay-per-request payments to prices falling toward marginal cost.
Don’t call me, I’ll text you. News and tutorials on Telegram, in Italian, as soon as they’re out.
Follow me on Telegram