Skip to content
Cameron Mills

All projects

A personal AI agent system, now used by a beta group of 10

The Life Tracker

Scheduled AI agents that run email and research, now used by a beta group of 10 selected users. Acting on their feedback, I ran refactoring rounds to simplify the codebase and widened Chief of Staff's predictions to cover more tasks. This page is the map, and each part has its own page.

Code on GitHub

The product story

User
Me first, and now a beta group of 10 selected users. Its first job is making sure nothing falls through the cracks: deadlines, admin and application cutoffs.
Problem
Obligations arrive in calendars, mailboxes and voice notes, and keeping a tracker up to date takes the time it is meant to save.
Options
An app I keep up by hand, or agents that do the upkeep. I set one rule, that it should take very little of my own time to keep running.
Decision
Agents do the upkeep inside a safety layer: no agent on the Mac can make a payment, and anything an agent may not do comes to me as a one-line command. I set the model-routing rules from an analysis of 519 agent runs by task type and model. Those are runs, not users.
Outcome
Used by a beta group of 10 selected users. Acting on their feedback, I ran refactoring rounds to simplify the codebase and widened Chief of Staff's predictions to cover more tasks. This page does not yet define active use, give a feedback example or show a measurement taken after those changes.
Next
Count how many of the 10 use it each week, and record one piece of feedback with the change it led to and what happened after.

Its parts

Dashed lines join each part to Life Tracker. Arrows run from the part that feeds to the part it feeds, labelled with what passes between them.

  • Life Tracker agent system: Running unattended AI agents reliably.
  • Polaris: One local dashboard where AI agents hand work back to me.
  • Dreaming: An overnight AI pass over a day's notes, capped and checked.
  • Perpetual Memory: Giving AI agents a memory that survives a cleared session.
  • Steward, the managed agent runner: Keeping long AI agent runs alive through usage limits and full contexts. Fed by Threadhaven (checkpoint).
  • /build skill: One build request in, a supervised multi-stage build out.
  • Life Tracker voice layer: Turning speech into filed work, with the source of every transcript on record.
  • LT Capture for iOS: An iPhone voice recorder that shows what the Mac did with each note.
  • Humaniser and writing-style system: A house style check that fails a rewrite which loses a number, name or citation.
  • Practice engine: Practice sessions that save every answer before the next question.
  • Threadhaven: A small checked task handover that Codex and Claude Code can both read. Feeds Steward, the managed agent runner (checkpoint).

How it works

The Life Tracker at a high level: what comes in, what it remembers, who does the work, what keeps it safe and what comes back to me. Select a layer to open it, and follow its link to that part's own page.

What it is for

The Life Tracker is my personal command centre, and its core is an AI agent rather than an app. Its first job is making sure nothing falls through the cracks: deadlines, admin and application cutoffs. I set it one rule, that it should take very little of my own time to keep running, so the agents do the upkeep.

How the parts connect

Work moves in one direction. Things come in, are remembered, are worked on by agents inside a safety layer, and come back to me as something I can act on. Each part has its own page.

Intake. It reads the places my work arrives: calendar feeds and my mailboxes. Voice notes are sorted by an agent into a reminder, a task or a note through the voice layer, and the LT Capture iPhone app shows what happened to each one.

Memory. Perpetual memory is a bank of short pages in which every claim quotes its source. Agents are handed the relevant pages before they start, and a handover routine lets a fresh session pick up where the last one stopped.

The agent system. Every unattended run goes through one agent manager, which pauses a run before the usage limit and resumes it afterwards. A supervisor only opens the next stage of a build once the previous stage's work actually exists. Steward feeds it from a timed queue, and the /build skill turns one request into a staged build.

Safety. A standing-orders file sets what each agent may do, and a hook stops any agent on the Mac from making a payment. Anything an agent is not allowed to do itself comes to me as a one-line command. Unattended runs send mail only after an adversarial review of each email, and reach git only through a Fable reviewer.

Outputs. Polaris is one window for everything waiting on me, from questions to answer to commands to run. A practice engine runs question-by-question sessions, and every agent that writes prose follows one house style.

Overnight. A chief of staff pass re-kicks failed jobs, closes clearly finished threads and writes what to do next. Dreaming reads the day's notes for links, duplicates and conflicts, and leaves a short report by morning. Its headline reaches the daily email. A morning scan re-checks every quotation in the memory bank, and any run paused at the usage limit picks up again after the reset.

Where it stands

It is now used by a beta group of 10 selected users. Acting on their feedback, I ran refactoring rounds to simplify the codebase and widened Chief of Staff's predictions to cover more tasks. I set its model-routing rules from an analysis of 519 agent runs, not users, by task type and model. It is in daily use, with a scheduler running whatever is due. In its first two weeks, 249 agent runs finished. Each part's page covers its own results and what did not work.

Stack

  • Python
  • Claude Code
  • Swift
  • iCloud Drive

How these figures are sourced

The outcome figure is agent-reported: it is taken from a test log or report written by the agents that built and ran the system. The status was read from installation status from commit c909757 on 1 Oct 2026, beta group from my CV of 28 Oct 2026. Every figure and claim on this page is tied to a source file or a CV line in the site's content.