Confucius Code Agent-Scalable Scaffolding for Large-Scale Repositories

The "model wars" are over—the real breakthrough is in the scaffolding. In this video, we dive into the Confucius Code Agent (CCA), a massive leap in AI software engineering that just hit a 54.3% Resolve Rate on SWE-Bench-Pro. The most shocking part? CCA (using Claude 4.5 Sonnet) actually outperformed Claude 4.5 Opus running on Anthropic’s own proprietary scaffold. This proves that how an agent structures its memory, orchestration, and tool use is more important than the raw size of the LLM itself. What makes the Confucius SDK different? The creators at Meta and Harvard didn't just build a better prompt; they built a system around three distinct axes: Agent Experience (AX), User Experience (UX), and Developer Experience (DX). Key Features We Cover: • The Orchestrator: A dual-interface system that supports both native JSON and XML-style tags to ensure compatibility across various models. • Hindsight Learning: How the Note-Taking Agent (F2) creates Markdown artifacts of failure modes so the agent never repeats the same mistake twice. • Hierarchical Memory: Using Adaptive Context Compression to handle massive repositories without hitting token limits. • The Meta-Agent (F4): A "master agent" that uses a Build-Test-Improve loop to automatically design and refine new agents from natural language specs. We also compare CCA’s "minimal intervention" style against Claude Code’s multi-agent delegation, explaining why subagents can sometimes lead to over-engineering and context loss