Agentic Engineering

Preface

Agentic Engineering is a term first coined by Simon Willison. It encompasses the application of software engineering principles to the use of LLM coding agents.

Patterns

Ask, Plan, Execute, Reflect (APER)

Taken from Agentic Engineering by Brendan O'Leary. Reflect phase added by me.

A pattern you have to apply yourself. Instructing the agent to follow this pattern will lead to worse results, as some of the context and reasoning will be lost for following this pattern instead of doing actual work.

1. Ask

Objective

• Understand the system
• Find the right files
• Don't plan yet; just gather information

Output

• Research file
  • Files involved
  • Line numbers for key functions
  • Data flow
  • Dependencies and side effects
  • Edge cases to consider

2. Plan

Objective

• Outline exact steps
• Include testing (how to verify that the goal is achieved)
• Be explicit about assumptions and uncertainties

Output

• Plan file
  • Step-by-step plan
  • Specific files and lines to edit
  • Test commands to run
  • Rollback strategy if things go wrong

3. Execute

Objective

• Execute the plan
• Keep context low (<= 100K tokens) (Smart Zone / Dumb Zone by Matt Pocock)
• Review each change
• Commit frequently

4. Reflect

Objective

• Review the process and outcome
• Identify what went well and what could be improved
• Document insights for future reference

Output

• Modified SKILLS.md, AGENTS.md, ...
  • If something went wrong, add a skill for how to avoid that in the future
  • If a skill did not do its job, update it to be more effective next time
  • If a new technique was discovered, add it to the relevant skill

Skills

Skills are a set of atomic markdown files that describe specific concepts, tasks, or techniques.

• t128n/skills — curated collection of skills sourced from various repositories
• mattpocock/skills — collection of skills by Matt Pocock
• skills.sh — CLI and browser for skill files

References

• Agentic Engineering by Simon Willison