Context
Engineering
for AI Agents
“Context Engineering is the discipline of designing, structuring, and managing the information environment that an AI agent reasons within — so it acts intelligently, reliably, and efficiently.”
Prompt Engineering told the model what to do. Context Engineering teaches it how to think — by precisely controlling everything that fills its context window: memories, tools, instructions, history, and live data. It is the craft behind every capable AI agent in 2025.
§ 01
The Three Pillars of Context
Pillar One
What Goes In
Everything placed into the context window before the model reasons. This is your engineering surface.
- System prompt & persona
- Retrieved memories
- Tool schemas & outputs
- Conversation history
- User input + files
- Background knowledge
Pillar Two
How It’s Structured
The arrangement, format, and priority of context elements. Order and structure dramatically affect model behavior.
- XML / Markdown delimiters
- Priority ordering (top = high weight)
- Compression & summarization
- Chunking strategies
- Token budget management
- Few-shot example placement
Pillar Three
How It Evolves
Context is dynamic. Good context engineering manages its lifecycle across turns, tasks, and time.
- Memory write / read loops
- Context window pruning
- Re-injection strategies
- State serialization
- Long-horizon planning context
- Forgetting & summarizing
§ 02
Anatomy of an Agent’s Context Window
Typical context window composition — a 200k token agent context
System Prompt
Long-Term Memory
Tool Schemas
Conv. History
User Input + RAG
Output Reserve
§ 03
Context Engineering Lifecycle
The Full Context Engineering Loop — From User Input to Agent Response
Memory Architecture — Four Types of Agent Memory
Context Pruning Strategy — Managing the Token Budget
“The bottleneck in agent performance is rarely the model. It is almost always the quality of information you put in front of it.”
— Core principle of Context Engineering
§ 04
Context Engineering in Practice
System Design
Customer Support Agent
Task: “I want to return the shoes I bought last week. They don’t fit.”
Engineered Context Window
[SYSTEM] You are a helpful support agent for ShoeStore. Be empathetic. Policy: 30-day free returns.
[MEMORY] User: Sarah Chen. Tier: Premium. Past issues: 2 resolved. Preferred size: US9.
[TOOLS] lookup_order(id), initiate_return(order_id, reason), check_inventory(sku)
[INPUT] “I want to return the shoes I bought last week…”
[MEMORY] User: Sarah Chen. Tier: Premium. Past issues: 2 resolved. Preferred size: US9.
[TOOLS] lookup_order(id), initiate_return(order_id, reason), check_inventory(sku)
[INPUT] “I want to return the shoes I bought last week…”
- 1System prompt sets persona + policy constraints
- 2Memory injects customer profile (no need to ask who they are)
- 3Agent calls
lookup_order→ injects result into context - 4Calls
initiate_returnwith correct order ID - 5Response + interaction saved back to memory store
Engineering
Autonomous Coding Agent
Task: “Add unit tests for the payment module and fix any bugs you find.”
Engineered Context Window
[SYSTEM] Senior Python engineer. Write pytest tests. Only edit files explicitly requested.
[MEMORY] Project: FastAPI app. DB: PostgreSQL. Pattern: Repository. Prior sessions: 3 bug fixes.
[FILES] payment.py (840 tokens), models.py (420 tokens) — retrieved via RAG
[TOOLS] read_file, write_file, run_tests, search_codebase
[MEMORY] Project: FastAPI app. DB: PostgreSQL. Pattern: Repository. Prior sessions: 3 bug fixes.
[FILES] payment.py (840 tokens), models.py (420 tokens) — retrieved via RAG
[TOOLS] read_file, write_file, run_tests, search_codebase
- 1RAG retrieves only relevant files (not entire codebase)
- 2Memory provides project conventions — no re-explaining needed
- 3Agent reads, writes tests, calls
run_tests - 4Test result injected back — agent iterates on failures
- 5Session summary written to long-term memory
Research
Deep Research Agent
Task: “Write a competitive analysis of EV battery manufacturers.”
Engineered Context Window
[SYSTEM] Expert analyst. Cite sources. Structure: Executive summary → Details → Conclusion.
[PLAN] Step 1: identify players. Step 2: gather data. Step 3: synthesize.
[SCRATCHPAD] [running notes & intermediate findings kept here, pruned as context fills]
[SOURCES] Retrieved articles injected progressively via RAG
[PLAN] Step 1: identify players. Step 2: gather data. Step 3: synthesize.
[SCRATCHPAD] [running notes & intermediate findings kept here, pruned as context fills]
[SOURCES] Retrieved articles injected progressively via RAG
- 1Plan injected upfront — agent follows structured reasoning
- 2Scratchpad in context stores intermediate work
- 3Sources retrieved progressively — old ones summarized & pruned
- 4Context budget managed: 60% sources, 30% reasoning, 10% output
- 5Final report written with full traceable citations
Personal AI
Long-Horizon Personal Assistant
Task: “Help me prep for my meeting with the investor tomorrow.”
Engineered Context Window
[SYSTEM] Personal assistant for Alex. Proactive. Anticipate needs. Concise.
[MEMORY] Investor: James Wong, Sequoia. Past meeting notes. Alex’s pitch deck v3. Alex’s goals: $2M seed.
[CALENDAR] Meeting: 10am, 45 mins. Location: Zoom. Retrieved from calendar tool.
[INPUT] “Help me prep for my meeting…”
[MEMORY] Investor: James Wong, Sequoia. Past meeting notes. Alex’s pitch deck v3. Alex’s goals: $2M seed.
[CALENDAR] Meeting: 10am, 45 mins. Location: Zoom. Retrieved from calendar tool.
[INPUT] “Help me prep for my meeting…”
- 1Rich memory profile means zero context re-establishment
- 2Calendar tool call injects live meeting details
- 3Past notes retrieved — agent knows investor preferences
- 4Agent produces tailored prep briefing in seconds
- 5Post-meeting: outcome saved to long-term memory
§ 05
Key Techniques Reference
| Technique | Category | What It Does | When To Use |
|---|---|---|---|
| RAG (Retrieval-Augmented Generation) | Context Filling | Fetches relevant documents from a vector DB and injects them into context at query time. Grounds the model in real, up-to-date knowledge. | Large knowledge bases, dynamic data, reducing hallucinations |
| Sliding Window | History Management | Keeps only the N most recent conversation turns in context. Older turns are dropped or summarized to make room for new input. | Long-running chatbots, multi-turn agents with token limits |
| Hierarchical Summarization | Compression | Progressively summarizes earlier parts of a conversation into increasingly dense summaries, preserving meaning while reducing tokens. | Long research sessions, hours-long agent tasks |
| Scratchpad / Chain-of-Thought | Reasoning Aid | Dedicated section in context for intermediate reasoning steps. Lets the model “think out loud” before committing to a final answer. | Complex multi-step tasks, planning, debugging |
| Few-Shot Examples | Behavior Shaping | Inject 2–5 high-quality input/output pairs into context to demonstrate desired format, tone, and reasoning style. | Structured outputs, specialized formats, consistent tone |
| Prompt Caching | Efficiency | Prefix parts of the context (e.g. system prompt + tools) for reuse across many calls. Dramatically reduces latency and cost. | High-volume applications, static system prompts |
| Tool Result Injection | Agentic Loop | After tool execution, the result is inserted back into context in a structured format so the model can reason about it in the next step. | All agentic tool-use scenarios |
| Semantic Memory Retrieval | Long-Term Memory | Stores past interactions as vector embeddings. At runtime, retrieves the most semantically similar past facts and injects them into context. | Personalized assistants, cross-session continuity |
| Priority-Based Ordering | Structure | Places the most critical instructions at the top of context (highest attention weight). Less important content goes later or is summarized. | All agent systems — always apply this principle |
| XML / Markdown Delimiters | Structure | Uses explicit tags like <system>, <memory>, <tools> to help the model distinguish between different types of context content. | Complex contexts with multiple distinct sections |
Bestseller #1
Bestseller #2
Bestseller #4
Bestseller #5
