Formation
Letting evidence write the principles
Opinion-based guidance gets ignored and evidence gives the guidance authority. So, I traced every principle back to a finding. I measured how instances of delegating overlapped with other patterns across 81 other scenarios. For example, prompt showed a 60% overlap, generate showed a 56% overlap, and create showed a 24% overlap. This provided direction to clarify points of friction, and redefine what pattern “delegate” multiple times..
Widen the definition
Agent-driven analysis narrowed down “delegate” applicable to agents only. However, human delegation such as, assigning an issue to a colleague through actionMenu or using Copilot to do so, is the same intent. I changed it to include both humans and agents and provide additional guidance which helped provide high-level guidance and cover all implementations.
Context is costly for agents; trading specificity for broader patterns to stay lean and fit new cases
Every fetch costs an agent context and there is a thin line between specificity and scenarios. Trimming down on specific examples like examples for “delegate entry points”, helped maintain high-level resonance and provide space for any new instance of delegate in future.
Recovery and exit from a delegation process
Two gaps: returning to a session later and any sort of graceful degradation from unexpected results. We call these error states and edge cases in UX. An external accessibility review pushed for a lot more work here. In particular, there is a grey area between accept and reject actions depending upon scenarios. After analysis and evaluation, the outcomes are now: accept, request changes, revert to a prior state, and try a different approach.
AI reviewers helped security measures
I saw the AI reviewer, running a degraded pass, flag that "Persist reasoning steps and tool calls" may encourage storing secrets, tokens, and internal prompts. Three humans had missed it. I decided to rewrite the guideline. Because a design system is the worst channel for a bad security default. It changed the page to a high-level activity log that avoids sensitive content.
