Bounded Context in Agents¶
Core Idea¶
An Agent does not operate on unlimited knowledge or infinite state.
Instead, it operates within a bounded context:
Context = What the Agent currently knows + can access + can reason over
This boundary is one of the most important constraints in real-world agent systems.
What is a Bounded Context?¶
A bounded context is:
The limited slice of information an Agent can use at any given moment to make decisions.
It includes:
- current conversation history
- retrieved documents (RAG)
- tool outputs
- memory snapshots
- system prompts
- environment state
Why “Bounded” Matters¶
Even though models feel powerful, they are still constrained by:
1. Context Window¶
Finite tokens → finite memory
Example:
8K / 32K / 128K tokens limit
The Agent cannot see everything at once.
2. Tool Latency¶
Not all information is immediately available.
Need to search → wait → observe
3. Relevance Filtering¶
More context ≠ better decisions.
Too much information causes:
- distraction
- noise
- hallucinated reasoning
- slower planning
Bounded Context vs Memory¶
They are related but not identical.
Memory¶
Persistent storage across time
Examples:
- vector database
- chat history
- long-term facts
Bounded Context¶
Active working set used right now
It is a subset of memory.
Agent Context Structure¶
A typical agent context looks like:
┌──────────────────────────────┐
│ Bounded Context │
├──────────────────────────────┤
│ System Instructions │
│ User Goal │
│ Conversation History │
│ Retrieved Knowledge (RAG) │
│ Tool Outputs │
│ Working Memory State │
└──────────────────────────────┘
Everything the Agent uses for reasoning comes from here.
Why Agents Need Context Management¶
Without context control:
Everything is thrown into prompt
This leads to:
- context overflow
- degraded reasoning
- high cost
- unstable behavior
Context as a Working Set¶
Think of bounded context like CPU memory:
Disk (Long-term memory)
↓
RAM (Bounded context)
↓
CPU (Reasoning step)
The Agent only operates on what is in "RAM".
Context Construction Pipeline¶
In real systems:
User Input
↓
Retrieve Memory
↓
Call Tools
↓
Fetch External Data
↓
Assemble Context
↓
LLM Reasoning
This assembly step is critical.
Context Selection Problem¶
The hardest problem is not generation.
It is:
What should be included in the context?
Too little context:
- missing information
- wrong decisions
Too much context:
- noise
- token overflow
- poor reasoning focus
Example: Coding Agent¶
Goal:
Fix failing test
Bounded context may include:
- test output
- relevant source file
- recent git diff
- previous attempts
But NOT:
- entire repository
- unrelated modules
- old logs
Example: Research Agent¶
Goal:
Write report on AI agents
Bounded context:
- top retrieved papers
- summaries of articles
- user constraints
Not included:
- all web pages ever visited
- irrelevant search results
Context Window vs Bounded Context¶
These are different:
Context Window¶
Hard model limit
Bounded Context¶
System design choice
You can have:
- large context window but poor selection
- small context window but excellent selection
Selection quality matters more.
Context Engineering (Modern Term)¶
Modern agent systems increasingly focus on:
Context Engineering > Prompt Engineering
It includes:
- retrieval strategy
- memory ranking
- summarization
- compression
- tool filtering
Context Compression¶
To fit within limits:
Summarization¶
Long history → short summary
Filtering¶
Keep only relevant info
Abstraction¶
Details → high-level representation
Failure Modes¶
1. Context Flooding¶
Too much irrelevant data:
Agent confused
2. Context Loss¶
Important information dropped:
Agent forgets constraints
3. Stale Context¶
Old information remains:
Agent uses outdated facts
Relation to Agent Loop¶
Bounded context is updated every loop:
Observe
↓
Update Context
↓
Reason
↓
Act
↓
New Observation
↓
Update Context
So context is not static—it evolves.
Key Insight¶
An Agent is not just a reasoning system.
It is a system that continuously constructs
and reconstructs its bounded context.
Design Principle¶
Good agent design:
Optimize context quality,
not context size.
Key Takeaways¶
Bounded context = active working state of an Agent
It is:
- dynamic
- limited
- continuously updated
Agent performance depends heavily on:
- what enters the context
- what is excluded
- how information is compressed
Context is not passive data.
It is the Agent’s working mind state.