Agents

The agents I run today: what they do, where they run, and the authority they’ve earned. Across the fleet, drafting is autonomous; acting is gated. Patterns explains the rules and their limits.

Last updated October 24, 2026.

Coach

Reads my watch and my training log every morning, drafts the day's prescription with a verification pass, and posts it to Discord and to Dispatch, my phone app; a button under the Discord brief, wired to my account alone, pushes the workout onto the watch. A slower second loop audits the program itself every 6 weeks, outcomes versus assumptions.

Status: live, daily.
Runs on: a frontier cloud model.
May: draft freely; write to the watch behind my tap.
May not: act on any device without the tap.
Last shaken down: August 2026.
Stories: part one, the agentic rebuild, the watch, the program audit, the 52 that never was, the couch-store detour
Patterns used: the house contract

Mailroom

Reads my Gmail every 15 minutes and files it under a handful of labels. A 27-billion-parameter local model makes every classification (it replaced the 14B and its home-trained adapter in August), and the system improves weekly from my corrections.

Status: live, every 15 minutes.
Runs on: a local 27B model on the homelab GPU; a cloud model sees only what the weekly improver is explicitly fed, and that now includes my draft-and-sent pairs.
May: label anything; archive bulk mail out of the inbox; draft replies, never send them, for the rare mail from a real person; hand tracking numbers it spots to the package tracker's relay (extraction is plain pattern-matching; a local model may propose a display name).
May not: send (the drafting path has no send call); delete anything a human wrote to me; beyond a label and its own unsent draft, a human's mail stays untouched.
Last shaken down: August 2026 (the 27B beat the adapter-tuned incumbent on the full corpus).
Stories: the guardrail, the red pen, the referee, the card, writing back
Patterns used: earned autonomy · propose, don't execute · graders need graders · the red pen

Columbo

Picks up what the alert rules forward: reads logs, runs read-only diagnostics over SSH, and posts a verdict with its evidence to Discord.

Status: live, on every forwarded alert.
Runs on: a frontier cloud model, from a separate machine (the fire-marshal rule).
May: investigate read-only; propose alerting-config changes, rendered literally.
May not: execute anything without my emoji approval; touch anything beyond alerting config.
Last shaken down: September 2026 (the chaos batches, on top of the quarterly drill).
Stories: the on-call engineer, letting it act, the drill, the test container, the filter regression
Patterns used: propose, don't execute · the fire marshal rule

Radar

Radar, the alert router, stood down in July when the noise it was hired to filter got fixed at the source. Then a new 27B model passed the exam the retirement left behind (on the recorded corpus it would skip 64% of investigations and none of the real faults), so the seat is live again.

Status: active in the routing seat (September 2026), with a quarter of its calls audited by the cloud model so a miss has somewhere to show up.
Runs on: a local 27B model on the homelab GPU.
Stories: the bake-off, the router that solved itself out of a job, the card, the test container

Quill

Quill, the ghostwriter, drafts every post on this blog against a voice specification my edits keep training (the colophon makes the full confession).

Status: live, per post.
Runs on: whichever frontier model currently has the seat.
May: draft, revise, run the pre-publish checks.
May not: publish; the draft flag is a door it cannot open.
Stories: the colophon
Patterns used: the red pen

Scout

My Saturday-night editor. It reads the week's decision logs and commit histories, pitches post candidates to Discord (however many clear the bar that week), and learns my taste from my thumb verdicts. A thumbs-up doesn't start a draft; it starts an interview, and the draft stays blocked until I've answered its questions about what actually happened. Scout finds and interviews the stories; Quill drafts and revises against the voice specification.

Status: live, weekly.
Runs on: a frontier cloud model for the pitches and drafts; a local 27B model on the homelab GPU tailors the interview questions.
May: pitch; interview me; write full drafts on a branch.
May not: publish; drafts wait behind the same flag Quill can't lift.
Stories: the story scout
Patterns used: silence isn't feedback

Patchday

Keeps the software on my homelab servers current. Once a week it finds which of the containers running our family's photos, media, and home automation have new releases (Renovate does the finding), reads each changelog, and posts a digest to Discord with a Skip button per update and one Install button for whatever I don't skip. A small executor holds the pipeline's only write path and applies exactly what I approved.

Status: live, weekly.
Runs on: a frontier cloud model for the changelog summaries; discovery and the executor are plain code.
May: install exactly what I approve, one tap for everything I didn't skip; majors never get a button, apart from a short list of stateless apps where reverting the image is the whole rollback.
May not: mutate anything unattended.
Last shaken down: July 2026 (the forced-failure drill).
Stories: Patchday, the reference model
Patterns used: the house contract · propose, don't execute

Rosie

Answers questions about the house and acts on it: reads the current state of every sensor, light, and thermostat, and takes plain-English requests in Discord ("is the garage closed", "warm up the office") through Home Assistant's API.

Status: live, in the house's Discord channel; an owner gate means only my messages are acted on.
Runs on: a local 27B model on the homelab GPU (since September 2026), handed the current house state with each message (reused for at most two minutes).
May: read every entity's state; eight control tools through Home Assistant's API (turn on, turn off, fan speed, brightness, thermostat, cover position, lock, unlock); an unlock waits for my Confirm tap, and thermostat setpoints stay between 60 and 80°F.
May not: create or edit automations or configuration; it has no tool that can.
Stories: Rosie joins the Johnstons
Patterns used: the house contract

Pepper

Pepper, the chief of staff, is the front door I am standing up for the whole fleet: one agent meant to take any plain-language request, in Discord, a chat tab, or by voice, and route it across the estate's tool servers for documents, notes, tasks and calendar, photos, and the homelab's own telemetry. The point is the timely details of life: when the flight leaves, what tomorrow looks like.

Status: in shakedown; it answers questions today, from Discord and from the phone, reaches the web through one hand-off per turn to a research worker that has none of its private tools, and no routine runs through it yet.
Runs on: a local 27B model on the homelab GPU.
May: read across the servers it fronts; propose task writes, one card and one tap each.
May not: execute a write without the tap; touch anything outside the servers it fronts.
Stories: Uptime was the wrong question, Pepper gets one phone call, Swiping away shouldn't cancel the answer
Patterns used: the house contract · structural absence beats in-loop refusal

Each agent appears here only after its story publishes. If this page and a post disagree, this page is current.