What the Claude Code Source Reveals About Agents

Haseeb Qureshi dug into the leaked Claude Code source and pulled out the engineering details that actually matter for anyone building coding agents. The terminal UI runs on React and Ink, not a traditional TUI framework. Compaction uses a four-layer fallback - proactive, reactive, snip, and context collapse with mid-conversation reversibility - which is far more sophisticated than just truncating context. Anthropic engineers get different system prompts than external users, with anti-hallucination guardrails and quality verification agents baked in. Permission denials route back to the model as tool errors rather than hard blocks, letting the agent adapt instead of failing. The real lesson from the leak isn't any single trick - it's that thousands of lines of infrastructure supporting four API calls represent years of operational learning that no model improvement alone can replicate.