Getting Stuff Done with Coding Agents — Using Them for Real Work, Well Beyond Writing Code

A lightning lesson on using coding agents — Claude Code, Codex — for real work, not just for writing code. Coding agents came to software first because software handed them a favorable environment: version control, tests, fast feedback, and an enormous amount of public training data. What you get out of the box is already a harness — a loop, tools, a shell, files, context, and knowledge. Your job is not to build that harness. It is to shape it around the work you actually do. Software gives you lessons, not a full playbook. Three things you have to adapt yourself: • Specifications — make the work explicit before the agent starts • Skills — teach the agent what it was never trained on: your knowledge, your preferences, your workflows • Verification — check the process, not just the output Then: how a project survives across many sessions, and why the workflow you build ends up acting as an operating system for the agent. I'm Stefan Jansen, founder of Applied AI and author of Machine Learning for Trading, whose companion code has ~20,000 GitHub stars. Over the last year I've shipped trading platforms and strategies, production ML pipelines, forecasting systems, AI agent systems, websites, and a 750-page technical book with coding agents. ————————————— WORKSHOP — Loop Engineering: Reliable Work From Coding Agents Saturday, July 18 · 4 hours · live on Maven https://maven.com/stefan-jansen/loop-... You start with Claude Code or Codex as the ready-made harness and leave having adapted it to your own work. We build a briefing agent together — it reads your feeds, drops duplicates, ranks what matters, and emails you one brief — then you adapt it to your own domain. Lightning lesson page: https://maven.com/p/8394ac/getting-st... ————————————— CHAPTERS 00:00 Introduction 02:26 What we'll cover 03:54 Coding agents grew up in software 05:12 A coding agent is a ready-made harness 07:18 Software gives lessons, not a full playbook 08:50 What you adapt: three things 10:02 Specifications — make the work explicit 12:48 Skills — teach what it wasn't trained on 14:40 Verification — check the process too 16:46 A project runs across many sessions 19:40 The workflow is an operating system 23:22 The workshop: build your environment 24:26 You shape the harness, not just the prompt