Routine, repeatable tasks are usually better as code – reliable, repeatable, verifiable – with AI steps embedded where needed, not handed wholesale to an agent. At the same time, fast development has revived the personal script: one-off, context-specific tools that were never worth building before, because the gap between "nothing" and "a maintained general-purpose tool" used to be too wide.
The key discipline is keeping the line clear between what you maintain and what you throw away. For me, every project and tool falls into one of these buckets:
- productionised - other users, uptime requirements, churn brings risk
- maintained - utilities and regular scripts that you need to keep in working condition but are free to amend, change and reset
- archive - never updated, but mined as starting points
- temporary one-offs
A key skill is identifying which bucket each tool you consider falls into, and applying only the necessary level of effort for that bucket.
Examples from our own work: a timesheet tool and a payroll-reconciliation script, both of which would have been agents if we'd been fashionable and worse if we had.
Key question/s: Which of your workflows deserve a tool rather than a chat window? What's in your personal toolkit?
Reading/links
- Building effective agents (Anthropic) – notable for how often it advises against agents
- Scripting with LLMs – Simon Willison's
llmtool – the embed-AI-in-a-script pattern, ready-made