# AI Agent Panels An AI Agent Panel is a group of [[AI Agents]] assembled for multi-angle evaluation of a specific input. Each agent in the panel evaluates through the lens of its own identity and expertise, then the results are aggregated into a unified scorecard. Panels are distinct from [[AI Agent Orchestration]] and [[Claude Code Agent Teams]]. Teams produce deliverables collaboratively. Panels evaluate existing work from multiple perspectives. The difference: teams build, panels review. ## Structure A panel definition specifies: - **Purpose**: when to convene this panel - **Agents**: a table listing each agent, their type (advisor, audience, or critic), and their role - **Input Required**: what the panel needs to evaluate - **Output Format**: how results are structured ## Agent Roles in Panels | Role Type | Purpose | Examples | |-----------|---------|---------| | **Advisor** | Domain expertise evaluation | Editor (structure), Marketer (conversion) | | **Audience** | Perspective of the target reader | Beginner (accessibility), Power User (depth) | | **Critic** | Adversarial stress-testing | Hater (logical holes), Skeptic (claims) | ## Orchestration Panel orchestration follows a sequential evaluation pattern: 1. Load the panel definition 2. Load each agent's identity (SOUL.md) 3. Evaluate the input through each agent's perspective, one at a time 4. Aggregate individual verdicts into a scorecard 5. Produce the top 3 fixes ranked by impact 6. Surface dissenting opinions if verdicts conflict Each agent gives a verdict (e.g., publish / needs work / kill it) plus 2-3 specific points. ## Benefits - Forces consideration of perspectives you'd naturally skip - Catches issues that a single reviewer would miss - Creates structured, actionable feedback rather than vague impressions - Composable: mix and match agents for different evaluation contexts ## Connection to Context Engineering Panels are an application of [[Context Engineering]] principles. Each agent evaluation loads only that agent's identity and the input to evaluate. This keeps context lean and focused, rather than trying to evaluate from all perspectives simultaneously (which would degrade quality). ## References - ## Related - [[AI Agents]] - [[AI Agent Identity]] - [[AI Agent Orchestration]] - [[AI Assistant Architecture]] - [[Claude Code Agent Teams]] - [[Context Engineering]] - [[Separation of Concerns]] - [[Composition over Inheritance]]