AI Agents for Beginners - 12. Context Engineering

AI Agents for Beginners - 12. Context Engineering

Source: microsoft/ai-agents-for-beginners - 12-context-engineering

What is Context Engineering?

Prompt Engineering vs Context Engineering

Types of Context

Context Types in AI Agents
Mermaid
%%{init: {'look': 'handDrawn'}}%%
flowchart TD
    Agent["AI Agent\nContext Window"]
    Agent --> I["Instructions\nPrompts, System Messages, Few-shot, Tool Descriptions"]
    Agent --> K["Knowledge\nRAG, Databases, Long-term Memory"]
    Agent --> T["Tools\nAPIs, MCP Servers, Function Results"]
    Agent --> CH["Conversation History\nMulti-turn Dialogue History"]
    Agent --> UP["User Preferences\nPreferences, Habits, Feedback"]

Strategies for Effective Context Engineering

1. Define Clear Results — Clearly define the desired outcomes of the task assigned to the AI Agent. You must answer the question: "What does the world look like when the agent is done?"
2. Map the Context — After defining results, determine "What information does the agent need to complete this task?" and map out where that information resides.
3. Create Context Pipelines — Once you know where the information is located, decide "How will the agent obtain this information?" Leverage RAG, MCP servers, and other tools.

Context Engineering Workflow
Mermaid
%%{init: {'look': 'handDrawn'}}%%
flowchart LR
    R["Define Clear Results\nClearly define outcomes"] --> M["Map the Context\nMap required info locations"]
    M --> P["Create Context Pipelines\nDeliver info via RAG, MCP, Tools"]
    P --> Agent["AI Agent\nEquipped with proper context"]

Planning Strategies

Managing Context

Context Management Strategies
Mermaid
%%{init: {'look': 'handDrawn'}}%%
flowchart TD
    CW["Context Window\n(Limited Size)"]
    CW -->|"Temporary single-session storage"| SP["Agent Scratchpad\nFiles / Runtime objects"]
    CW -->|"Long-term cross-session retention"| MEM["Memories\nSummaries, Preferences, Feedback"]
    CW -->|"When approaching limits"| COMP["Compressing Context\nSummarization / Trimming"]
    CW -->|"Complex parallel tasks"| MAS["Multi-Agent Systems\nDistributed independent contexts"]
    CW -->|"Code / Large-scale processing"| SBX["Sandbox Environments\nReturn results only"]
    CW -->|"Multi-step workflows"| RSO["Runtime State Objects\nSubtask result containers"]

Example of Context Engineering

• Checks available dates in the calendar in real time
• Recalls preferred airline, budget, and direct flight preference from long-term memory
• Identifies the tools needed to book flights and hotels

Common Context Failures

Context Poisoning

Context Poisoning Mitigation
Mermaid
%%{init: {'look': 'handDrawn'}}%%
flowchart TD
    LLM["LLM Response\n(Includes flight info)"] --> V{"Real-time API Verification"}
    V -->|"Verification successful"| CTX["Add to context\nProceed normally"]
    V -->|"Verification failed\nHallucination detected"| Q["Quarantine\nBlock context"]
    Q --> Fresh["Fresh context thread\nPrevent contamination"]

Context Distraction

Context Distraction Summarization
Mermaid
%%{init: {'look': 'handDrawn'}}%%
flowchart LR
    Long["Long Conversation History\n(Includes irrelevant content)"] -->|"Threshold exceeded"| Sum["Context Summarization\nCurrent goals / relevant details only"]
    Sum --> Next["Next LLM Call\nUse compressed summary"]
    Next --> Focused["Focused Response\nTailored to current request"]

Context Confusion

Context Confusion RAG Filtering
Mermaid
%%{init: {'look': 'handDrawn'}}%%
flowchart TD
    Query["User Query\n'How to get around Paris?'"] --> RAG["RAG over Tool Descriptions\nSelect relevant tools from vector DB"]
    RAG --> Focused["Focused Tool Set\nrent_car, public_transport_info"]
    Focused --> LLM["LLM\nInvoke appropriate tools"]
    LLM --> Response["Accurate Response"]

Context Clash

Context Clash Resolution
Mermaid
%%{init: {'look': 'handDrawn'}}%%
flowchart TD
    Old["Previous Instruction\n'Economy class'"] --> Conflict{"Conflict Detected\nNew instruction arrives"}
    New["New Instruction\n'Business class'"] --> Conflict
    Conflict -->|"Context Pruning"| Prune["Remove previous instruction\nKeep new instruction only"]
    Conflict -->|"Scratchpad"| Scratch["Reconcile in Scratchpad\nDetermine final consistent instruction"]
    Prune & Scratch --> Clean["Consistent Context\nAgent operates normally"]

Chat History Reduction with Agent Scratchpad

• Token counts grow linearly with each user message and assistant response
• Resending the full conversation history every time causes prompt tokens to account for the majority of costs
• Eventually, if the conversation exceeds the context window, the model truncates it or returns an error

Creating a Context-Aware Agent

context_aware_agent.py
python

Simulating a Long Conversation

long_conversation.py
python

Context Growth in Long Conversations
Mermaid
%%{init: {'look': 'handDrawn'}}%%
flowchart TD
    T1["Turn 1\nJapan, Sushi, Temples, Photography"] --> T2["Turn 2\nBudget $3000, Solo, 10 days, April"]
    T2 --> T3["Turn 3\nTest context retention"]
    T3 --> T4["Turn 4\nAccommodation: Traditional Japanese Ryokan"]
    T4 --> T5["Turn 5\nDate change: October autumn foliage"]
    T5 --> T6["Turn 6\nRequest full summary"]
    T1 & T2 & T4 & T5 -->|"Accumulated context"| CTX["Context Window\n(Growing)"]
    CTX -->|"Short conversation: Normal"| T6

Context Summarization Pattern

1. The agent identifies key facts from the conversation
2. Calls the summarization tool to store them persistently
3. Because the summary captures core essentials, older messages can be safely pruned

summarization_tool.py
python

summarization_demo.py
python

Summarization and Scratchpad Pattern
Mermaid
%%{init: {'look': 'handDrawn'}}%%
flowchart TD
    Conv["Multi-turn Conversation\nAccumulated preferences, budget, dates, interests"] --> Agent["SummarizingTravelAgent"]
    Agent -->|"Identify key facts"| Tool["summarize_preferences()\nStore compact summary"]
    Tool --> Sum["[SUMMARY] User preferences recorded: ..."]
    Sum -->|"Older messages can be pruned"| Trim["Context Window Reduction"]
    Trim --> Next["Next LLM Call\nSummary-based response"]
    Next -->|"Continuous continuity"| User["User\nResponse retaining previous context"]

vacation_agent_scratchpad.md
markdown

Summary

Process Flowchart
Mermaid
%%{init: {'look': 'handDrawn'}}%%
flowchart LR
    Root["Context\nEngineering"]
    Root --> PE["vs Prompt Eng.\nStatic vs Dynamic"]
    Root --> Types["Types of Context\n5 Sources"]
    Root --> Plan["Planning\nStrategies"]
    Root --> Mgmt["Managing\nContext\n6 Strategies"]
    Root --> Fail["Common\nFailures"]
    Types --> I["Instructions"]
    Types --> K["Knowledge"]
    Types --> T["Tools"]
    Types --> CH["Conversation\nHistory"]
    Types --> UP["User Preferences"]
    Plan --> P1["Define Clear Results"]
    Plan --> P2["Map the Context"]
    Plan --> P3["Create Pipelines"]
    Mgmt --> M1["Scratchpad"]
    Mgmt --> M2["Memories"]
    Mgmt --> M3["Compression"]
    Mgmt --> M4["Multi-Agent"]
    Mgmt --> M5["Sandbox"]
    Mgmt --> M6["Runtime State"]
    Fail --> F1["Poisoning\nHallucination propagation"]
    Fail --> F2["Distraction\nOver-focus on history"]
    Fail --> F3["Confusion\nTool overload"]
    Fail --> F4["Clash\nInformation conflict"]

• Context Engineering goes beyond Prompt Engineering to build a repeatable and reliable process ensuring agents are equipped with the dynamic information they need over time
• Context is composed of five types: Instructions, Knowledge, Tools, Conversation History, and User Preferences
• Effectively managed through Planning Strategies (Define Results → Map Context → Create Pipelines) and six practical strategies (Scratchpad, Memories, Compression, Multi-Agent, Sandbox, Runtime State)
• Recognize four Context Failures—Context Poisoning, Distraction, Confusion, and Clash—and mitigate them via verification, summarization, RAG, and pruning
• Maintain context continuity while reducing token costs even across extended conversations using the Chat Summarization + Agent Scratchpad pattern