TKTomas Kruty

Builds · Morning brief · 27 September 2026

How the morning brief runs

Nine small agents and three scripts, one email, one mp3. What each one does between 05:00 and 07:00, and why one big agent was the wrong shape.

Every morning at 05:00 UTC a cloud routine clones one repository and reads one line: read the orchestrator prompt and follow it exactly. Two hours later there is one email in my inbox with the eight to twenty things that matter to my work, and a three-minute spoken version in four voices.

The shape

A collector script reads the feeds, the searches, Hacker News, arXiv and a few scrapers in the background while two agents triage my inbox and read my calendar. Then the curator, the only agent that ever reads the raw items, walks them section by section and writes a shortlist with a question per item. Three or four researchers spend a fetch budget answering those questions from the primary sources. Four section writers turn the answers into the email's beats. A podcast writer turns the sections into a script for a host, an analyst, a mentor and a newsreader. An assembler joins it all, an editor fixes what it finds, and a sender delivers the email with the mp3 attached.

Why nine agents and not one

The first version was one agent holding the whole day. Past a certain length it drifted, forgot the meme, and nothing errored. Nine small agents with state in files, and three plain scripts for collecting, assembling and sending, cost less and never lose the plot. Every step writes one file; if an agent fails, the run writes one line to a problems file that ends up in the email footer, and continues. A brief with a footnote beats no brief.

Teaching it

A one-sentence reply to the email becomes a habit the next morning already follows. The reply is turned into a positively stated rule, every writing prompt renders the rules, and the editor checks the brief against them before it is sent.