Agent Progressive Disclosure & Eval Framework — March 5, 2026

Changes Made

1. Progressive Disclosure (Anthropic pattern)

All agent AGENTS.md files restructured:

• Level 1: SOUL.md (identity, always loaded) — unchanged
• Level 2: AGENTS.md (core workflow, <70 lines each) — slimmed down
• Level 3: references/ (project context, loaded on demand)

2. Shared Project Context Files

• ~/clawd/agents/references/medschools-context.md — Stack, schema, deploy, gotchas
• ~/clawd/agents/references/hedge-context.md — Architecture, data providers, safety rules
• Agents load ONLY the relevant context file when working on a project

3. AGENTS.md Line Count Reduction

• atlas: 90 → 51 lines
• kai: 93 → 48 lines
• kevin: 129 → 43 lines
• liz: 101 → 41 lines
• sage: 109 → 68 lines
• maya: 227 → 50 lines
• designer: 56 → 56 lines (already lean)

4. Eval Framework

3 sample tasks per agent in ~/clawd/agents/evals/<agent>.json

• Kai: responsive component, dark-theme table, mobile overflow fix
• Atlas: RLS policy, FastAPI endpoint, query optimization
• Kevin: feature brief, sprint planning, bug triage
• Liz: strategy brief, backtest review, incident response
• Sage: security code review, competitive research, performance review
• Maya: SEO blog outline, content calendar, landing page copy
• Sam: component spec, dashboard layout, design review

Each eval has specific grading criteria (pass/fail).
Run monthly or after significant agent config changes.