Decisions from real projects, with the measurements that produced them. No cadence promised — these go up when something is worth writing down.
Beside a session for each project, I now run sessions for roles: one drafts my posts, one watches the chats, one looks after the feel of my games, one develops photos. Why a narrow session beats one agent that does everything, what keeps them organised, and what it costs.
My coding agents keep working while I sleep. Not because they never stop, but because a manager agent watches them, answers what the plan already answers, and leaves the rest for my morning. Four nights, measured, and what went wrong.
Work in progress. My manager agent and its overview are moving into the Claude desktop app, for people who never open a terminal: only what needs me, most urgent first, beside the chat. What it looks like so far, what the app taught me, and what is still wrong.
My manager agent kept the records, but I still had to dig through its scrollback to find what needed me. So it got a pane beside the chat: the one thing to do next and why, which sessions are waiting on me, and the approvals only I can give.
Five coding agents were never the bottleneck. Coordinating them was. So I gave them a manager that keeps the records, writes the specs and never relays an approval, and a fence so that sessions which must never talk cannot.
I built a dictation tool for myself in a day. The interesting part was not the app — it was discovering that the on-device model beat the larger one on speed and accuracy, and that the naming told me the opposite.