Building AI Agents with Claude Code — Session 4: Parallel Agents & Feedback Loops

What You’ll Build

A system that generates competing variants in parallel, evaluates them adversarially, and feeds results back to improve future runs.

Key Patterns

• Complexity triage — classify before processing, skip overkill for simple tasks
• Adversarial evaluation — evaluators that attack, not praise
• Uncorrelated perspectives — independent reviewers catch what solo reviewers miss
• Fresh context via files — file-based state over LLM memory for reliability
• Parallel generation — non-determinism becomes an advantage with evaluation
• Feedback loops — agents that learn from past results
• Workflow DAGs — composing everything into pipelines