"Context Engineering — What to Show the Model (Harness Series 2/6)"

Second of the seven engineering problems from Series 1/6. Deciding what to show the model. Good context → good answer. Break the context, and even GPT-5.5 starts answering like GPT-2.

Part 1 introduced the Agent = Model + Harness thesis. This article covers the most consequential axis of that harness: Context Engineering.

Context Engineering = "the discipline of designing systems that dynamically assemble the right information for each step, rather than loading everything and hoping."

This guide walks through the principles, techniques, and per-tool implementations (Claude Code, Cursor, oMLX, etc.).

Series Roadmap (6 parts)

  1. What Is Harness Engineering? — definition, history, why
  2. Context Engineering ← this article
  3. Memory Systems — session/long-term, vector vs SQL
  4. Tools & Sandboxing — tool spec, MCP, VM isolation
  5. Multi-Provider Routing — cost/quality routing, fallback
  6. Evaluation & Ops — eval harness, observability, runner

1. The Context Window Is a Shared Budget

One token ≈ 0.75 English words. A 128K window holds about 96K words. A 1M window holds roughly 750K words — five novels.

That budget is shared across: - System prompt - Retrieved documents (RAG) - Conversation history - The model's own output - Tool call results (MCP and friends)

Key principle: "Bigger context window = better outcome" is wrong. What you spend the budget on is what matters.

Major-Model Context (April 2026)

Model Context
GPT-5.5 1.05M tokens (2× rate above 272K input)
Claude Opus 4.7 200K (1M optional)
DeepSeek V4 1M tokens
Kimi K2.6 256K tokens
Qwen3.6-27B 262K (extensible to 1M via YaRN)

A million tokens is now standard — but filling it is rarely the answer.

2. Why "Don't Fill It" — Lost in the Middle

Long-context research consistently shows:

Models use the beginning and the end of context best. The middle gets lost (the "Lost in the Middle" problem).

Filling 1M tokens means: - Higher TTFT (prefill is O(N²)) - Higher cost - Lower accuracy (information buried in the middle) - Reasoning burnout: a 50-step workflow at 20K tokens/call = 1M cumulative — output quality degrades

Bottom line: Context should be small and precise.

3. Five Core Techniques

3-1. Auto-Compact

When context utilization nears the cap, compact older content into a summary.

Claude Code: at roughly 98% utilization, runs auto-compact. Older turns get summarized to free space.

oMLX takes a different angle — external persistence via SSD KV cache rather than in-window compaction.

3-2. Hierarchical Processing

Convert older context into successively more abstract representations: - Raw: "User edited line 42 of file X to handle case Y..." - Level 1: "User edited auth logic in file X" - Level 2: "Session focused on auth + routing"

Deeper compression = more density, more detail loss. Find the balance.

3-3. Tool-Output Limiting

Tool calls eat context fast.

Claude Code default: MCP tool output capped at 25,000 tokens, with a warning at 10,000. Servers may opt into up to 500,000 characters, persisted to disk instead of held in context.

Don't load full tool results into the window. Big results go to disk; the model sees a summary or index.

3-4. Reference vs Inline

Instead of dropping a whole code file into context: - Put just the file path in the system prompt - Let the model use a Read tool when needed

This makes context-filling dynamic. A 100K file isn't carried every turn.

3-5. CLAUDE.md / Project Context Files

Persistent project information lives in a separate file that's auto-loaded each session.

Claude Code: auto-loads CLAUDE.md from the project root every session. Custom commands, coding conventions, forbidden patterns.

This file is part of Guides (per Part 1's classification). It accumulates without volatilizing.

4. Claude Code's 5-Layer Compaction Pipeline

Per arXiv 2604.14228 ("Dive into Claude Code"), context management is a 5-layer pipeline:

  1. Live conversation buffer — most recent turns intact
  2. Recent tool results — partial retention
  3. Compacted history — older turns summarized
  4. Project context — CLAUDE.md + project metadata
  5. System primitives — core system prompt

Each layer has its own compression ratio and lifetime. When layer 1 fills, content evicts to 2~3; once those fill, content gets summarized.

vs Plain FIFO

Most chatbots just delete the oldest message. Claude Code summarizes and keeps. Information loss is minimized while space is reclaimed.

5. MCP — Standardized Tool Interface

Anthropic's Model Context Protocol (MCP), released November 2024, is the standard for tool-call wiring. As of 2026, Claude Code, Cursor, and OpenAI Codex CLI all use MCP servers.

Context Implications

  • MCP servers expose external systems (DB, API, FS) as tools
  • Tool results land in the context window → pressure
  • The Claude Code 25K/500K limits exist to manage this pressure

Good vs Bad MCP Servers

  • Good: returns summarized results; large data exposed via IDs to fetch separately
  • Bad: dumps a 50K-row SQL result back, blowing up context

Tool authors share responsibility for context engineering.

6. Production Patterns

6-1. System-Prompt Diet

  • Bad: "You are an extremely helpful, friendly, cooperative, deeply thoughtful, accurate..."
  • Good: "Goal-driven. State results directly. No filler."

Trade adjective stacks for behavior commands. Saves tokens. Gives the model an actionable directive.

6-2. Externalize In-Progress State

Big tasks save mid-step state to external files (plan.md, tasks/) → reload after a context clear.

CLAUDE.md pattern: every task ends with a snapshot in tasks/sessions/. Auto-loaded next session.

6-3. Sub-Agent Isolation

Long searches and analyses go to a sub-agent. Main context only sees the result, not the search noise.

Claude Code's Agent tool implements this pattern.

6-4. RAG vs Long Context

  • RAG: 100K-doc index, fetch 5K per query
  • Long Context: load all 100K in-window

Comparison: | Item | RAG | Long Context | |---|---|---| | Cost | Low | High (full prefill) | | Accuracy | Depends on retrieval | Self-attention | | Update | Re-index only | Re-fill every time | | Use when | Structured retrieval | One-shot whole-doc reasoning |

Even with 1M-token models, RAG retains a clear cost edge.

7. Tool-by-Tool Comparison

Tool Auto-compact Tool-output cap Project context External memory
Claude Code ✅ (~98%) 25K/500K CLAUDE.md ✅ (Skills, MCP)
Cursor 3 ✅ — .cursorrules ✅ (cloud agent)
OpenClaw ✅ — CLAUDE.md (compat) ✅ (memcore)
Hermes ✅ — own ✅ (FTS5 + vector)
oMLX — (model-level) — — SSD KV cache

oMLX persists KV cache at the inference server layer (not the model). Repeated calls with the same system prompt make prefill effectively free.

8. Anti-Patterns (the common mistakes)

8-1. "Just stuff it into a big-window model"

A 1M window doesn't help if you fill it: - TTFT explodes (prefill ∝ N²) - Cost explodes (GPT-5.5 above 272K input doubles) - Middle information lost (Lost in the Middle)

8-2. "Pipe tool results directly into context"

A 50K-row SQL dump pollutes the context window for every subsequent turn. Summarize and externalize.

8-3. "Cram everything into CLAUDE.md"

A 30K-token CLAUDE.md eats 30K from every session start. Keep core only; move infrequent knowledge to docs/ and load on demand.

8-4. "Preserve all conversation history"

Long conversations need compression and externalization. The most common failure when transitioning from chatbot patterns to agents.

Bottom Line

Context Engineering is not a window-size race.

Principle One line
Small and precise 5K accurate beats 1M filled
Dynamic assembly Build at runtime, not statically
Externalize Memory, tools, files live outside — pulled in when needed
Auto-compact Compress before the cap
Layer separation System / project / session / tool results have different lifetimes

The single takeaway: "The context window is a shared budget. Every token has a price."

Part 3 (next) covers preserving information outside the window — Memory Systems.

First-Party Sources

  • "Dive into Claude Code": arxiv.org/abs/2604.14228 (Liu et al., 2026-04-14)
  • Anthropic MCP spec: modelcontextprotocol.io
  • Lost in the Middle: arXiv:2307.03172
  • Martin Fowler harness article: martinfowler.com/articles/harness-engineering.html

๋Œ“๊ธ€

๋Œ“๊ธ€ ์“ฐ๊ธฐ

์ด ๋ธ”๋กœ๊ทธ์˜ ์ธ๊ธฐ ๊ฒŒ์‹œ๋ฌผ

Agent Memory Engine (2/10) — Building an AI Agent Memory System with SQLite Alone

← 1/10 SQLite Memory Engine f… ๐Ÿ“š Series Index 3/10 memcore Completed → A memory engine that runs without daemons, without dependencies, anywhere ํ•ต์‹ฌ ์š”์•ฝ This post covers how to design an AI agent memory engine (memcore) on a single-file SQLite backend. - Structure: 25 modules, ~6,464 lines, 22 DB tables - Principles: no daemon, single-file portability, no required external dependencies, host-agnostic - Core techniques: 4-layer hybrid search (topic · vector · FTS5 · LIKE), U-tag dialectic-based confidence evolution, ingestion-layer bias rejection rules What This Post Covers One approach to building AI agent memory without a server or cloud vector DB. Specifically: (1) the limitations of text-file-based memory, (2) vector DB alternatives compared and selection criteria, (3) how to layer search tiers, (4) how to evolve confidence scores over time, and (5) how to block biased memory accumulation at the ingestion stage. Design Principles: Why Single-File SQLite Text-file-based ...
Read more »

"ML Foundations (9/9) — PyTorch vs TensorFlow, and the Road to Local LLMs"

์ด๋ฏธ์ง€
← 8/9 Deep Learning Architec… ๐Ÿ“š Series Index (series end) The final part. So far we've focused on models. Now we focus on the tools that actually run them . We'll lay out the philosophical difference between PyTorch and TensorFlow, and trace how Hugging Face transformers , llama.cpp , MLX , and Ollama built the bridge to running large language models on your own machine. By the end you should have the full mental model of "download a pretrained LLM and serve it locally." 0. Learning Objectives Explain the eager-vs-graph difference between PyTorch and TensorFlow. Explain, in graph terms, how autograd automates the backward pass. Use Hugging Face transformers ' AutoModel , AutoTokenizer , and pipeline abstractions. Describe llama.cpp's GGUF quantization, INT4 inference, and CPU-first flow. Describe Apple MLX's use of unified memory and how it differs from PyTorch. Run a local LLM with Ollama and call it through an OpenAI-compatible API. ...
Read more »

"RAG Core Study (14/26) — Evaluation Sets with RAGAS & DeepEval"

์ด๋ฏธ์ง€
Series overview: Series index ๐Ÿ“š Series Index Without an evaluation set, every RAG improvement is a story, not evidence. RAG systems fail in more than one place: retrieval, context selection, answer generation, and grounding. That means "looks better" is not a metric. Part 14 explains how to build a golden dataset , how RAGAS and DeepEval fit into the workflow, and why teams must separate retrieval quality from answer quality if they want tuning decisions to hold up. 0. Prerequisites Part 13 reranking — system quality now depends on multi-stage ranking. Part 1 grounding — a correct-looking answer is not enough. Part 12 Hybrid — multiple retrieval settings need comparison on fixed data. 1. Learning Objectives Build a small but useful golden dataset for RAG. Distinguish retrieval metrics from answer metrics . Use RAGAS and DeepEval in the right roles. Avoid the common traps in judge-model-based evaluation. 2. ํ•ต์‹ฌ ์š”์•ฝ An evaluation set for RAG is ...
Read more »

"ML Foundations (8/9) — Deep Learning Architectures: CNN, RNN, Attention"

์ด๋ฏธ์ง€
← 7/9 Deep Learning Training ๐Ÿ“š Series Index 9/9 PyTorch vs TensorFlow → What did we keep adding on top of an MLP? Vision went CNN, sequences went RNN/LSTM, and eventually both converged on Attention and the Transformer. Each architecture is easier to remember if you read it as what it gave up to gain something else . This part is the one-line summary of the decisive inflection points: LeNet → AlexNet → ResNet → LSTM → Attention → Transformer. 0. Learning Objectives Explain why CNN's convolution, pooling, and weight sharing match images so well. Trace what got unblocked from LeNet → AlexNet → VGG → ResNet as depth grew. State BPTT for RNNs and the vanishing/exploding gradient problem. Write the LSTM gate equations (forget, input, output) with intuition. Write the attention formula in query/key/value form and the scaled dot-product variant. State the precise reasons Transformers replaced RNNs (parallelism + long-range dependencies). 1. ํ•ต์‹ฌ ์š”์•ฝ CNN : weight sharin...
Read more »

"ML Foundations (7/9) — Deep Learning Training: Optimizers, Regularization, Initialization"

← 6/9 Neural Networks ๐Ÿ“š Series Index 8/9 Deep Learning Architec… → Building an MLP, as we did in Part 6, is not the same as getting it to train well . The right optimizer, regularization, initialization, and learning rate — when these line up, deep networks converge. When they don't, the network refuses to learn at all. This part is the catalog of those crafts, with formulas, paper citations, and code in one place. 0. Learning Objectives Compare and write the update rules for SGD, Momentum, Nesterov, and Adam. Explain how Dropout, BatchNorm, and LayerNorm work and where they belong in a model. Derive the variance formulas for Xavier and He initialization and match them to activations. Implement step, cosine, and warmup learning-rate schedules in PyTorch. Explain why gradient clipping is effectively required for RNNs and Transformers. Diagnose the most common training failures (NaN, plateau, overfitting) and apply first-line fixes. 1. ํ•ต์‹ฌ ์š”์•ฝ SGD : \(w \leftarro...
Read more »

OpenClaw to Hermes Migration (2/13) — What to Preserve, Partially Port, or Discard

← 1/13 Current Structure Inve… ๐Ÿ“š Series Index 3/13 What Moves to Hermes → A code review record classifying operational scripts into three tiers: preserve, partial-preserve, and discard. What This Post Covers An asset classification framework (preserve / partial-preserve / discard) for migrating a legacy agent system to a successor A concrete case applying the Strangler Fig + Subset Migration strategy to a memory pipeline Classification results evaluated across LOC, external dependencies, and replaceability axes A method for slimming down a bloated single file (1,310 LOC) using the migration as a natural entry point Problem Definition: Port Everything or Cut Deep? The OpenClaw inventory comprises multiple agents, multiple scheduler/daemon processes, and multiple operational scripts. When moving to the successor system, Hermes, the choice is not binary. A full port carries accumulated technical debt along with it; a minimal port discards hard-won domain logic. Code-revie...
Read more »

AI Agents I Built (5/7) — Building an Automated Blogger API Publishing System

← 4/7 My Life Organizer Agent ๐Ÿ“š Series Index 6/7 Dual → Designing a Markdown-Based Auto-Publish Pipeline with Blogger API v3 + OAuth 2.0 Key Summary Blogger API v3 + OAuth 2.0 enables a pipeline that converts Markdown files into publishable HTML posts blogger_publish.py handles the full flow — frontmatter parsing → body cleanup → HTML conversion → CSS injection → API post — in a single script For bulk publishing, quota management (delay, retry, TOC suppression) is the critical design concern Platform Selection Criteria API access is mandatory for AI agent-driven content publishing. A comparison of platforms: Platform API Cost AdSense Decision WordPress.com REST API Paid plan required Paid only Cost overhead Ghost REST API Self-hosted or paid Manual setup Ops overhead Blogger v3 API Free Native integration Adopted Blogger is free, exposes a REST API, and integrates natively with AdSense. For personal blog automation, this combination i...
Read more »