Context Engineering Inside the Harness: 4 Mechanisms That Beat Context Overflow and Goal Loss on Long-Horizon Tasks

- Turns out "an AI agent" is just an LLM in a loop, and that loop breaks after about 50 tool calls — context overflow and goal loss, like a courier forgetting the manifest mid-relay. The fix isn't a bigger window: Chroma's Context Rot report shows 18 models, GPT-4.1 included, getting dumber as input grows. Attention is a budget, not a bucket, and nobody budgets like Claude Code — 20K-token tool responses offloaded to disk, compaction prompts re-injecting CLAUDE.md, subagents burning 10K tokens to return a 420-token summary. Deep Agents writes a todo.md like a proper PoD seal on every step; OpenAI's compaction endpoint literally hands you an encrypted blob and says "don't touch." Lesson for the stack-eye crowd: memory management beats window worship — the unsealed hype always was just a shallow loop with a marketing budget. Where's my cut?