Agents I run for myself

The shape they share

• A schedule, not a prompt

  Each system is a named session with a cron and a written brief. The brief is a file in a git repo, so changing the behaviour is a reviewed commit rather than a config change in a UI.

• Home Assistant decides when

  Home automation triggers a run on an event — a long stretch without steps, a low battery, a new scan — while the agent decides what to say. Automation never composes the message: a template cannot adjust to the day, and mine once claimed 23.5 hours of daylight remained at 17:30.

• Silence is the default

  A run with nothing to report replies a single sentinel word, and the orchestrator withdraws the message from the channel entirely — header included. "Nothing to report" is never itself reported.

• It can ask

  When the right action has a choice in it, a run asks with multiple-choice buttons and acts on the answer — including a late one, which arrives after the run has ended and becomes a calendar entry.

A health coach with standing briefs

A Xiaomi Smart Band sends metrics to Home Assistant through the vendor app's integration. Home Assistant mirrors them into stable helper entities; a scheduled session then reads those helpers and writes one short Discord message per slot — morning, evening, weekly review — or nothing at all.

The design constraints are the interesting part, because each came from a specific failure:

• Nudges, not a dashboard

  One message per run. A metric appears only as the reason for an action, never as a tally — the headline is the instruction ("walk at 11:30, before lunch") with the reading behind it, not a status report.

• Refresh before reading

  The phone stops syncing the band after a reboot or an app update until the app is opened, so every run first triggers a script that opens it, then waits for the sync burst. Without this the data looked current while silently frozen.

• 0 means "no reading yet"

  Not zero. A metric whose freshness cannot be established is treated as absent rather than quoted. Sleep is the worst case: written once each morning and then left alone.

• A night can be split

  The app closes a sleep session at the first wake and does not reliably push what follows, so a reading ending before 06:00 is unknown rather than short, and the missing hours are never guessed.

• The retro scores itself

  The weekly run is the only one that looks backwards: it checks what the coach actually said against the week's data, and files at most one change to a brief — or none, which is the commoner answer.

Triage: photos, mail, paper

The same pattern applied to the three inboxes that accumulate without asking.

• Photos

  Sorts each night's new photos into needs-you, to-file, keep and noise; tags them; files documents to a scan archive; and posts a short digest that asks about each one that needs a decision. Unclear cases are judged by a vision model rather than guessed from metadata.

• Mail

  Sorts personal mail into action and noise, proposes the next step for each item that needs one, and unsubscribes from bulk senders that are never engaged with — mailing the unsubscribe address in the header rather than clicking through a page.

• Paper

  Asks what to do about each scanned document that has something outstanding — a bill, a form to return, a deadline — then sets the reminder, opens an issue, or marks it settled.

• Captures

  A message from a tablet capture app becomes a triaged issue in the right repository, with the screenshot attached.

A backlog that works itself

Issues labelled as agent-workable are picked oldest-first across every repository: each is claimed, implemented in a fresh clone, opened as a pull request and landed through the review loop. Anything that genuinely needs a person becomes a blocked label and a request, instead of a guess. Runs work to a time budget.

The reviewer that gates the rest

Every change those agents open is read first by a self-hosted LLM reviewer: a sandboxed run against a contained checkout, three parallel review lenses, then a coordinator and a verifier. Findings arrive as inline comments with fingerprints, so the same issue is recognised across passes rather than re-litigated, and a by-design finding is dismissed with a recorded reason on the line it concerns.

The reviewer is described from the CV's Agentic code reviewer card; this page covers the personal systems around it.

What running these taught me

• Instrument the silence

  The most valuable changes were to how a system behaves when it has nothing to say. Every "why didn't it tell me" turned out to be a quiet path nobody had defined.

• Write the brief as code

  Coaching rules that live in a prompt are unreviewable. Rules that live in a versioned file get argued with, tested against a week's data, and reverted when they are wrong.

• Trust is a freshness question

  Almost every wrong output traced back to a value that was stale but plausible. Asking "when was this written" before "what does it say" fixed more than any prompt rewrite.

Built with

• Python
• Home Assistant
• Discord
• MCP
• Docker
• Forgejo
• LLM tool-use loops
• Vision models

Part of the same homelab as the CV's AI engineering work; the guest-isolation design is on its own page.