AI agent skills that verify their own work.
Each agent task includes a checklist for what "done" means. Before the task reports success, it asks a shared checker to verify every item. No external orchestration required.
The problem
They declare success when they are 80% done. A skeleton-project skill would scaffold a folder, push to GitLab, and declare done. Half the time the remote push had silently failed. The folder existed. The commit existed. The remote did not.
Self-Enforcing Workflows are the fix.
External enforcement
"Someone will check your work before it ships."
Self-enforcement
"You check your own work before you hand it off."
Explore
The Pattern
A completion checklist inside every task, a shared checker that reads it, and a meta-skill that repairs bad checklists — all held together by a 3-strike enforcement loop.
Explore the pattern →The Ecosystem
Orchestrators, sub-skills, quality gates, and meta-skills — each with a defined place in the quality chain, enforced at write time by an integrity hook.
Explore the ecosystem →Prior Art & Novelty
How SEW compares to CrewAI guardrails, Design by Contract, CI/CD quality gates, and what appears to be genuinely new.
Explore prior art →