Life OS: the agent fleet and second brain I run my own week on.

Life OS is the personal AI operating system I built for myself: a markdown second brain, a local knowledge engine (gbrain) that reads and searches it, and a set of scheduled agents that run parts of my week without me opening a laptop. It's not a product — I'm not selling it — it's the proof that I use what I sell. Pre-revenue, private, and still being built.

The problem

I run several things at once — locum shifts, HMR work, client builds, a few products in progress, a personal brand. The context for all of it used to live in scattered notes, chat threads, and my own memory, which meant I re-explained the same background to myself constantly and lost track of open loops. I didn't want a SaaS tool with my life locked inside someone else's database. I wanted files I own, on disk, that both I and AI agents could read.

What I built

The base layer is a markdown vault — one file per topic, per venture, per decision — that I edit directly. Sitting on top of it is gbrain, a local knowledge engine that indexes and searches that vault so an agent can pull the right page instead of me pasting context by hand. On top of that, scheduled agents handle recurring work: a morning brief that pulls together what's live across my ventures, and agents that draft or triage specific recurring tasks. I also run two AI vendors side by side — Claude and OpenAI's Codex — under a written protocol that says which one owns a task and which one only reviews it, so they don't collide or duplicate work.

Where the human gates sit

Nothing here writes to the vault, sends anything externally, or takes an irreversible action without me. The vault itself has a single-writer rule — one agent holds the lease to edit the canonical files at a time, specifically so two agents can't clobber each other's edits. Every agent-to-agent handoff produces a written brief and a return I check before I trust it, rather than just accepting "done." The protocol exists because the failure mode isn't dramatic — it's quiet drift, one agent acting on stale context the other agent already corrected. Writing that rule down, and only trusting a return when there's actual evidence behind it, was the fix.

What broke or was hard

The knowledge engine has needed real maintenance, not a one-time setup — its local database has broken across upgrades more than once and needed a manual fix each time. Getting two different AI vendors to hand off work cleanly took an actual written protocol before it worked reliably; ad hoc prompting between them just produced duplicated or contradicted work. And a system like this only stays honest if I keep the vault current — the day I stop updating it, every agent built on top of it is reasoning from stale information. That's a discipline problem more than a technical one, and it's the part I trust least right now.

Current status

As of 27 August 2026: Life OS is private, running daily, and not for sale. I'm not publishing time-saved figures or automation-count metrics — I don't have numbers I'd stand behind, so I'm not inventing any. What I can say plainly is that scheduled agents genuinely run parts of my week now, the vault is the thing I actually consult before making a decision, and the Claude/Codex protocol is a real document I follow, not an aspiration. It keeps changing as I find the next weak point.

This is also the reason I hold a firm line with clients: anything with a similar failure mode — silent drift, an agent acting on data nobody checked — gets a human gate before it goes near a business that isn't mine. I write more about that discipline on the build log.