Managing several inboxes, calendars and task sources across Boyle the Builder, Dundee Founders Collective and personal life left little room for the higher-value work. The hypothesis: my AI build experience would transfer directly to running an always-on personal agent.
I stood up a dedicated, always-on agent, named Dave, on his own machine, with access to Gmail, calendar and Telegram, following a published build guide. The plan was inbox triage, urgent-item flagging, and event digests into my Airtable database.
It failed, and I will say so plainly. Forty pounds of credits vanished in days while Dave still knew too little about me to make a single useful call. I decommissioned him. The hypothesis was wrong: experience building AI systems does not substitute for the thing an agent actually runs on.
An agent is only as good as its governed, persistent context. Before autonomy is worth anything, an agent needs an operating contract: what it may do, what it must never do, and the context it carries between sessions. I had built none of that, so Dave burned credits learning nothing. That diagnosis became the design principle behind the knowledge system I now run everything on, which is the cheapest £40 lesson I have ever bought. I may bring Dave back one day. Contract first, next time.