# context-engineering

A Tsinghua Shenzhen team proposes NLAH (Natural-Language Agent Harnesses): moving agent control logic from code into structured natural language, executed by an IHR runtime. Experiments show harnesses can reshape agent behavior patterns entirely, but more structure doesn't always mean better results. Dan McAteer argues harness engineering matters as much as model capability.

Tired of tweaking prompts and swapping models, only to find your AI agents still can't 'evolve'? This post reveals a deceptively simple secret: a Markdown-based context system that turned one person's agents from clumsy interns into autonomous powerhouses in just 40 days — using the exact same model throughout.

The core message is simple: most people don't fail because the model is weak — they fail because their context management is a mess. The author advocates starting with a minimal CLI workflow and iterating with rules, skills, and clear task endpoints. It's not about chasing new tools; it's about making your agent's behavior controllable, verifiable, and convergent.

A Context Engineer at Sully.ai built his entire digital brain inside a Git repo: 80+ markdown/YAML/JSONL files, no database, no vector store. Three-layer Progressive Disclosure, Episodic Memory, and auto-loading Skills — so the AI already knows who he is, how he writes, and what he's working on the moment it boots up.

An AI Agent burns 34,500 tokens of system prompt every single conversation turn. The author used layered loading (always-on vs on-demand) plus a dual-model strategy to cut monthly costs from $568 down to $120-150 — a 75% reduction. Full breakdown with real numbers inside.

The 'Context Tax' in AI brings triple penalties: cost, latency, & reduced intelligence. Nicolas Bustamante's 13 Fintool techniques cut agent token bills by up to 90%. A real-money guide for optimizing AI context, covering KV cache, append-only context, & 200K token pricing.

Junde Wu from Oxford + NUS got fed up with coding agents forgetting everything between sessions. So he built OneContext — a Git-inspired context management system using file system + Git + knowledge graphs. Works across sessions, devices, and different agents (Claude Code / Codex). The underlying GCC paper achieves 48% on SWE-Bench-Lite, beating 26 systems. Backed by an ACL 2025 main conference long paper.

For vibe note-taking to work well, you must force Claude Code to be 'picky.' Use a 4-layer filtering mechanism (file tree → YAML descriptions → outline → full content) to make it more selective. This pattern is called Progressive Disclosure.