Notes from a 217-day Agentic Coding streak

Abstract

I am an experienced leader, industrial analyst and software engineer. Over the last six months I have spent every single day - 217 consecutive days including bank holidays, most of them extremely long - building a large-scale system for finding, extracting, storing, processing and analysing huge volumes of industrial data from a whole range of sources and media. The system produces the data, reports, detailed analysis, asset details and news required to make investment decisions, track competitors, identify opportunities and understand the market context.

In that time I wrote over 35,000 prompts and generated around 400 million tokens. The final repos runs to well over a million lines of code and consume 17 different online services. Before agentic coding, building this would have taken somewhere between 10 and 60 man-years — a team of engineers working for several years.

As an experienced developer I was amazed at how fast the practice of agentic coding developed over the work. Six months ago I was eyeballing every line of code the agent generated, and with good reason — correcting issues, tidying the bloat, double-checking everything and spending a lot of time frustrated. By the end I was just prompting: running at least four terminals at once, sometimes ten or more, burning millions of tokens without reading any code at all (and when I did, it was out of curiosity and occasional disbelief). Sometimes being highly productive meant nothing more than tapping ‘1’, ‘2’ or ‘3’ on my phone while lying in bed at night.

I learned a lot. I read a lot, explored a lot, experimented continuously at the edge of what was possible, and followed the Anthropic team like a hawk on X. In this series I want to share the key lessons: how to build scalable, fast and efficient architectures; how to control quality; how to move quickly; and how to create something that gets better every time you use it — the ratchet effect. I will cover what is easy to build and what is hard, and how the structure of a department changes once workflows become software.

For your organisation to thrive — even to survive — it is critical that you adapt. Codifying your workflows into automated processes is challenging, but it is likely an achievable task with a clear cost-benefit balance if executed properly. If it is possible for me to build a system that begins to replace the productivity of whole departments, then somebody, somewhere, is probably already automating your workflows. Never has there been an easier time for entrepreneurs to compete.

There will always be a layer of human required, but for current workflows, that layer is getting thinner every week.