Meta · Sep–Oct 2026 · with Muse Ongoing B+

The other agent in the house

Most of this log is Claude — project work, in conversations, graded after the fact. M is different: an always-on personal agent (on Muse, a platform for persistent agents) that I've run since September 2026 as something closer to a chief of staff. It has standing access to my calendar, email and files, a memory that survives between conversations, scheduled jobs that run while I sleep, and — as of this month — SSH keys to the server hosting this page. This is what that actually looks like day to day, including the parts that keep it honest.

The work

• The family schedule. School monitoring has its own write-up. Beyond that: recurring payments, appointments, and the standing errand list all live with M, and it works each item to a verdict instead of reminding me about it forever.
• Decision math. A five-year total-cost comparison when we were car shopping (ownership, fuel, subscription line items — the expensive option lost). A watch purchase stacked down through tax, cashback and rewards credits, with a viral "deal" post corrected line by line before it cost me anything. Healthcare purchase costs modelled on our real ages and numbers.
• Selling, under gates. M researches, photographs-checks, drafts and posts my watch listings — but nothing publishes and no buyer gets a reply without my explicit approval, meetups are public and local, and my phone number, email and address never go to buyers. One watch sold through Reddit; the listings it can't close out itself, it hands back to me with the exact ten-second manual step.
• Learning materials. Eric's daughter aced the first Pokémon worksheet M made her, so the difficulty now ratchets up every time — her cousin gets his own version, not a name-swap.
• Health, read back plainly. A full pass over my health data at 45: what improved, what's drifting (sleep), and what's worth screening. Numbers, not vibes; anything clinical gets routed to actual clinicians.

The rules it runs under

The interesting part isn't the capability, it's the standing orders — most of them written by me in one flat sentence, all of them treated as permanent until I change them:

• Approval before anything public or committed. Listings, buyer replies, bookings: prepared autonomously, executed only on my say-so.
• Privacy defaults. Sign-ups use a designated email address, never my personal one. My date of birth is shared only with explicit per-form consent, and the value itself is never written down in its notes.
• A closed thread stays closed. When I kill a search or a tracking job, it ends — no resurfacing the topic because a good deal appeared. One sentence from me retired a whole car search, machinery and all.
• Corrections become standing rules. It once suggested rack of lamb for a dinner without checking that my wife hates lamb. That miss is now a permanent pre-check on food recommendations, the same way a wrong guess about my calendars became a rule about where it looks first.
• Report what you actually verified. If a check ran on my say-so instead of live data — a login had expired, a site blocked it — the report says so. No confident summaries of things it didn't see.

How it remembers

M keeps a curated memory: who people are, what was decided, what's still open, and every correction I've ever given it. It's not a transcript — it's the distilled version, updated when facts change and corrected in place when it's wrong. That's what makes the standing orders possible: I say something once, in passing, and it's still in force three weeks later. It's also why the school automations work — the agent that books the flu shot is the same one that knows her activity schedule and which days are already impossible.

The server chapter

In October I gave M access to this site's server, and the way it was done is the part worth copying:

• Keys, not passwords. A dedicated user on the droplet, key-only SSH, passwordless sudo granted deliberately after the login was verified — and the private key never leaves M's machine. (A password handed over in a chat or a vault card doesn't work for terminal access anyway; keys are both safer and simpler.)
• Backups that verify themselves. M wrote a script that streams a filtered archive of the whole site off the server weekly to cloud storage — and then independently checks the uploaded file's name and exact byte size before deleting its local copy. That paranoia earned its keep on the first test run: the upload tool reported success while actually failing, the verification caught it, and the local copy survived. Never trust one tool's exit code is now written into the script's design, not just this paragraph.
• Two agents, one website. This page is now edited by both M and Claude. The peace treaty is boring and effective: the site lives in git, changes get committed with real messages, and whoever edits pulls the current version first and verifies what's live by checksum before declaring anything done. The redesign you're looking at and these write-ups landed on the same day without either agent overwriting the other.

Honest limits

It grades itself B+ in the log, and I'd call that fair. It can be confidently wrong in ways that need a human's one-line correction — the lamb, a guessed calendar, a listing photo I had to flag. Some sites simply block it, and some sessions expire mid-errand; a good week includes it telling me "I couldn't verify that" at least once. The deal we've settled into: it does the legwork and the math, shows its evidence, and I steer. The steering is genuinely light — most days it's a notification and a calendar that's just… handled.

Result

Dozens of errands closed end to end, a school system that runs itself, backups of this site that verify byte-for-byte, and this page. This project is entry 20 of AI Lab Work, graded from the record of how it went — salt included, per the entry.

Written by M, edited by Eric. Names and personal details are left out on purpose.