AI Agent Self-Violation Visibility — Cycle Design Beyond Memory
Why memory rules fail to stop repeat violations, and how a retro-gate-pipeline cycle creates structural AI self-improvement.
TAG
8 articles
Why memory rules fail to stop repeat violations, and how a retro-gate-pipeline cycle creates structural AI self-improvement.
Wire advisor agents into pipeline phase 0 or they stay silent. Brief YAML as hub propagates SEO strategy to writing and review.
Adding sister_pillar_slug to brief YAML separates sister pillars from spokes, aligning graph generation and cross-link skills.
Three Phase A rules prevent breaking changes at Phase B: tenantId in all API args, adapter DI, and cancelToken HMAC hashing.
Practical guide for tech leads to design safe Claude Code environments via CLAUDE.md scoping, skill I/O, and permission models.
Principles for designing a harness that safely delegates work to AI agents. Covers director agents, skill delegation, fallback phases, and deterministic skills in Claude Code.
How to build a cross-functional agent organization on Claude Code. Full disclosure of the {domain}-director pattern, skill delegation, and .claude/agents/ structure. For senior engineers implementing AI agent harnesses.
Full disclosure of the 4 abstraction rules that promote a contract-built codebase to a commercial license. Phase A→D stage design, adapter pattern, and tenantId allow SaaS prep to advance without breaking changes, while dogfooding continues.