Agentic Workflows: Engineering Principles for Non-Engineers

Background

I’ve been building non-coding agentic workflows for my own work (conference management, content production, research) and pairing with colleagues to build theirs. At the Paris Anthropic meetup, I presented on RAG to marketers and sales folks and got flooded with questions about agentic workflows. The questions were all about pains people had with ChatGPT, Claude, OpenAI: lost context, unreliable outputs, workflows that break after a few sessions. Pains I didn’t have, because I’d been applying engineering principles to non-engineering work.

That gap is what this research addresses.

Table of contents

  1. What is an LLM — statelessness, non-determinism, degradation, truncation
  2. From chat to agent — agents, tools, the LLM as a helper
  3. A real workflow — conference management little demo
  4. Getting started — start from pain, first principle: only build workflows around things you already do well
  5. One agent, one job — second principle: one agent, one job (Unix philosophy)
  6. Self-containment — third principle: self-containment (encapsulation)
  7. Reviewing changes — fourth principle: always verify non-deterministic output (git)
  8. One session, one task — fifth principle: scope sessions to a task, not to time
  9. Risks — sixth principle: least privilege, give the agent only the access it needs and can reliably handle
  10. Memory — tattoos in the folder, and why Claude Code’s automatic memory is an anti-pattern
  11. Skills and scripts — seventh principle: reduce the surface where non-determinism can act (attack surface reduction)
  12. Using third-party skills — eighth principle: practise supply-chain hygiene
  13. Hooks and linters — automating mechanical checks
  14. Tokens and costs — billing models, why long sessions cost disproportionately more, and watching the meter
  15. Change one thing at a time — ninth principle: change one thing at a time (OVAT, bisection), the maintenance principle
  16. Watching the cost — the status line as an always-on cost meter, and ccusage for daily and monthly totals
  17. Analysing your own work — the strategic payoff: contained work becomes a complete, queryable record, for self-assessment, job analysis, and delegation

Status

Early draft, chapters 1-17 published. More chapters planned.