In a medieval scriptorium, monks copied manuscripts by hand, one painstaking letter at a time. A single book could consume a year. Then the press arrived, and the labour of a year became the labour of a week. The monks did not vanish; the work simply moved up the ladder, from copying to editing, illuminating and judgement.
We stand at a similar threshold. AI agents are the new scribes: tireless, fast, and occasionally prone to inventing a saint that never existed.
What an Agent Really Is
Strip away the marketing and an agent is a loop: a model reads a goal, chooses a tool, observes the result, and repeats until finished or stopped. The power is in the tools, whether files, a browser, a shell or an API, and the danger is in the same place.
An agent is only as trustworthy as the walls around it.
Keep the Model on Your Own Hardware
There is a quiet convergence between this topic and the rest of my work. Open-weight models now run on consumer hardware, which means the model can run on your own machine, reading your own documents, without shipping your private ledger to a stranger. Local inference brings the local-first ethos to AI.
- Privacy: sensitive material never leaves the device.
- Resilience: the tool works when the API is down or the price changes.
- Sovereignty: the model you rely on cannot be quietly swapped beneath you.
Rules for Employing Automata
- Least privilege. Give the agent the narrowest set of tools that does the job.
- Reversible by default. Prefer actions that can be undone; require a human seal for those that cannot.
- Leave a paper trail. Log every action in plain text you can read.
- Verify, then trust. Treat output as a draft from an apprentice, not a decree from an oracle.
Where Judgement Stays
The scribe copies; the editor decides what is worth keeping. The future of work with AI belongs to those who keep taste, judgement and responsibility for themselves, and delegate only the copying.