Until now, work happened when you asked. Today you write standing orders: tasks that run on a schedule, unattended, and report to your Telegram. You'll build one from scratch — then adopt another from our lab's library of 70 working loops.
Learn the four parts of an agentic loop, put one real loop into service (a morning briefing to your Telegram), then pick and adapt a second loop from the Loop Library.
You need: Hermes + the Telegram bridge from Day 6 · the loops article read · 90 minutes.
The pipeline is you driving. A loop is the agent driving itself — for one defined job, inside fences you set. Every loop, from a morning reminder to this course's own homework-checker, has the same four parts:
The four parts of every agentic loop. Click each card to check yourself against last night's article.
Loops that change or delete files get extra suspicion: keep them in report-only mode for a week before letting them act.
Write the standing order — and dry-run it
I want a scheduled task called "morning briefing".
Task: read AGENTS.md and spec.md in my my-tool folder, then send me a Telegram message with three parts: 1) where the project stands (2 lines), 2) the single next step, 3) one question I should answer today.
Guardrails: read-only — this task may not modify any file. If the folder is missing or unreadable, report that instead.
Do a DRY RUN now: produce the message here so I can check the format. Don't schedule anything yet.Step done when: the dry-run message reads well. Tune the format now, while it's cheap.
Set the delivery address and the schedule
Scheduled results need a home. In your bot's Telegram chat, send /sethome — that marks this chat as the delivery channel. Then, back in Hermes:
The dry run looks right. Schedule "morning briefing" for every day at 07:30, delivered to my Telegram home channel. Also schedule a TEST run 3 minutes from now. Then show me how to list and cancel my scheduled tasks./sethome again in the right chat.✦ The test message arrived without a request. The trigger fired, the task ran inside its fences, the report came to you. You specified once — the work now repeats without you. That is agentic engineering.
Step done when: the unprompted test message lands in your Telegram.
Open the Loop Library — 70 working loops in five categories: Engineering (32 — code and project upkeep), Evaluation (16 — quality checks on schedule), Operations (9 — reminders, reports, tidy-ups), Content (7 — digests and drafts), Design (6). Every card has the full prompt, when to use it, verification steps, and a copy button.
Shop for your loop
Browse for 10 minutes with one question in mind: “what do I re-do by hand every day or every week?” Open 3–4 cards that catch your eye and read their Use when lines. Pick ONE whose job you actually need — not the coolest one, the one you'd keep.
Step done when: you've picked a loop and copied its full text (the “Copy full loop” button).
Adapt it — the four-line conversion
A library loop is a template; yours needs your folders, your schedule, your delivery. Paste it to Hermes with the conversion instruction:
Here is a loop from our Loop Library: [paste the full loop]
Adapt it to me:
1. TRIGGER: run it [your schedule — e.g. every weekday at 18:00].
2. TASK: keep the job, but point it at [your real folder / data].
3. GUARDRAILS: it may touch only [name the folders]; if it changes files, run report-only for the first week.
4. REPORT: deliver results to my Telegram home channel, every run.
Show me the adapted version, then do a dry run. Don't schedule until I say so.Step done when: the dry run produces output you'd want to receive on schedule.
Commission it — and know your kill switch
If the dry run is right: “Schedule it.” If the loop only reads — it's in service now. If it writes — leave it report-only this week (guardrail rule 3). Last thing, ask Hermes: “Show me how to pause or delete this loop.” Every standing order you create, you must also know how to cancel — that's what makes it engineering instead of magic.
✦ Two loops now work for you on schedule: one you built, one you adopted. Your mornings are staffed.
Step done when: both loops are scheduled and you've seen the pause/cancel commands.
Let both loops prove themselves, and share your adopted loop so the cohort can steal it.
homework complete — mark the day done below
Tick when the homework above is finished — it syncs to your account.
The trigger. A loop runs because the schedule fired, not because you asked. You specified once, set the fences, and the work repeats without you — the shift from using an agent to engineering with agents.
Bad, always. Rule 3: report every run, including “nothing to do”. A silent loop is indistinguishable from a dead one — and dead loops get discovered exactly when you needed them.
Because a library loop already survived contact with reality: its task wording, checks, and fences were debugged by someone else. You spend your effort on the adaptation — your folders, your schedule — which is the part only you can do. Invention is the upgrade, not the entry point.