The basic maths is starting to look embarrassing for traditional software development timelines. Claude Code, Anthropic's terminal-based AI coding agent, can now be configured to run autonomously for 24 hours or longer, working through backlogs, writing functions, running tests, and fixing its own errors while the person who set it going gets some sleep. According to Anthropic's own usage data, developers using Claude Code report completing in hours what previously took days. For a small business owner who needs custom software but can't justify a full-time developer on the payroll, that shift is not incremental. It's structural.
The Towards Data Science guide sets out a practical architecture for long-running agent sessions: structured task queues, checkpoint saves so work isn't lost if a session drops, and clear instruction sets that let the agent handle ambiguity without stopping to ask questions every ten minutes. The key insight is that most AI coding agents fail over long runs not because they run out of capability, but because they hit an unexpected decision point and stall. Solving that, through better upfront prompting and fallback logic, is what separates a 20-minute session from a 24-hour one.
This matters far beyond software development. The same agent architecture applies to data processing, report generation, content pipelines, and administrative automation. A healthcare practice administrator who needs patient correspondence templates built out across dozens of scenarios, or a school administrator processing data into structured formats, can point a Claude Code agent at the problem and return to a finished output. Research from McKinsey Global Institute estimates that knowledge workers spend roughly 60 per cent of their time on tasks that are, in principle, automatable. Long-running agents start to put a dent in that number in a way that one-off prompts simply don't.
For Scottish SMEs specifically, the cost angle is direct. Edinburgh developer day rates typically sit between £400 and £700 depending on specialism, according to figures published by Scottish Enterprise's digital skills reports. A Claude Code subscription costs a fraction of that monthly. The gap between what an AI agent can now do unattended and what a human contractor charges to do attended work is widening fast. That doesn't mean developers are redundant, it means a founder who understands how to direct these agents can now move like a team of five, not a team of one.
The practical barrier remains setup confidence. Most SME owners aren't software engineers, and the Towards Data Science guide is written with engineers in mind. But the underlying tools are becoming accessible: Anthropic's Claude interface, paired with clear task documentation and some basic terminal familiarity, is enough to get started. Business Gateway Scotland runs regular digital skills workshops, and the Scottish Government's Digital Growth Fund has historically supported businesses investing in exactly this kind of productivity infrastructure. The tools exist. The funding pathways exist. The only thing that doesn't scale automatically is the willingness to start.