← Back to Capabilities
👁️

Advisor Agent — Session Overseer

A second AI watching the first AI work
P1Core Pattern
Summary
An advisor agent that monitors and evaluates the primary Claude session. Automatically receives full conversation history, acting as a real-time overseer for the working session. Designed to improve qualitative output by providing independent assessment.
Architecture Diagram
Advisor Pattern — Separate Executor and EvaluatorWorker AgentExecutes taskAdvisor AgentEvaluates qualityresultfeedbackIterate until quality threshold met
Technical Details
This is architecturally significant: it's not just delegation (subagents doing work) but supervision (an agent watching another agent's quality). Pattern: worker agent does the task, advisor agent evaluates the work, improvements feed back into the next iteration. This dual-agent quality pattern — separate executor and evaluator — is a production-grade approach to AI reliability.
Implementation Pattern
TypeScript (conceptual)
// Advisor pattern: separate executor and evaluator
async function advisedExecution(task: Task) {
  // Worker does the task
  const result = await workerAgent.execute(task);

  // Advisor evaluates (receives full conversation history)
  const evaluation = await advisorAgent.evaluate({
    task,
    result,
    conversationHistory: getFullHistory(),
  });

  if (evaluation.quality < THRESHOLD) {
    // Feed improvements back for iteration
    return advisedExecution(
      applyFeedback(task, evaluation.suggestions)
    );
  }
  return result;
}
Architecture Insight
This is architecturally significant because it separates execution from evaluation — the same principle behind code review, QA testing, and peer review. The advisor doesn't do the work; it judges the work.
Official / Public Basis
Advisor agent pattern found in source architecture. Separate from subagent delegation — this is supervision, not task decomposition.
Governance Concerns
Dual-agent quality pattern adds cost but improves reliability. Must balance overhead vs. quality improvement. Not every task needs an advisor — reserve for high-stakes operations.
LightHope Ecosystem Mapping
LightHope — m8002 (quality control) as an advisor agent pattern, automated code review, content quality assurance, teaching quality monitoring