AI Agents with Context

Complete guide to context management – token budgeting, optimization strategies, multi-agent policies, and monitoring.

AI Agents with Context

See also: Five-Layer Agent Stack — Context is Layer 2.

PraisonAI provides industry-leading context management with smart defaults, lazy loading, and 6 optimization strategies.

Diagram

Feature	PraisonAI	LangChain	CrewAI	Agno
Smart Defaults	✅	❌	❌	❌
Lazy Loading (0ms)	✅	❌	❌	❌
6 Strategies	✅	❌	❌	❌
Per-Tool Budgets	✅	❌	❌	❌
Session Deduplication	✅	❌	❌	⚠️
LLM Summarization	✅	⚠️	❌	❌

Quick Start

Enable Context

from praisonaiagents import Agent

# Enable with defaults (auto-enabled when tools present)
agent = Agent(
    instructions="You are helpful",
    context=True
)

Custom Config

from praisonaiagents import Agent, ManagerConfig

agent = Agent(
    instructions="You are helpful",
    context=ManagerConfig(
        auto_compact=True,
        compact_threshold=0.8,
        strategy="smart",
        llm_summarize=True,
    )
)

YAML Config

context:
  auto_compact: true
  compact_threshold: 0.8
  strategy: smart
  llm_summarize: true
  tool_limits:
    tavily_search: 2000

What is Context?

Context is everything sent to the LLM in a single API call. It includes:

Diagram

System Prompt

Agent instructions, role, and goals (~2K tokens)

Chat History

User/assistant messages (variable)

Tool Schemas

Function definitions (~2K tokens)

Tool Outputs

Results from tool calls (~20K tokens)

Memory/RAG

Retrieved context (~4K tokens)

Output Reserve

Space for LLM response (~8-16K tokens)

How Context Flows

Single Agent Flow

Diagram

Multi-Agent Flow

Diagram

Optimization Strategies

When context exceeds the threshold (default 80%), the optimizer kicks in:

Diagram

Strategy	How It Works	Best For
`truncate`	Remove oldest messages	Simple chatbots
`sliding_window`	Keep N recent messages	Long conversations
`prune_tools`	Truncate old tool outputs	Tool-heavy agents
`summarize`	LLM summarizes old context	Critical context
`smart`	Combines all strategies	Production use
`non_destructive`	Tag for exclusion (undo-able)	Audit trails

Smart Strategy Flow

Diagram

Overflow Handling

Diagram

Level	Usage	Action
Normal	< 70%	No action
Warning	70-80%	Monitor
Critical	80-90%	Auto-compact triggers
Emergency	90-95%	Aggressive optimization
Overflow	> 95%	Emergency truncation

Token Budgeting

The Context Budgeter allocates tokens across segments:

Diagram

python

from praisonaiagents import ContextBudgeter

budgeter = ContextBudgeter(model="gpt-4o-mini")
budget = budgeter.allocate()

print(f"Model limit: {budget.model_limit:,}")      # 128,000
print(f"Output reserve: {budget.output_reserve:,}") # 16,384
print(f"Usable: {budget.usable:,}")                 # 111,616

Model Limits

Model	Context	Output Reserve
gpt-4o	128K	16K
gpt-4o-mini	128K	16K
claude-3-opus	200K	8K
gemini-1.5-pro	2M	8K

Per-Tool Budgets

Set different limits for different tools:

Diagram

python

from praisonaiagents import Agent, ManagerConfig

agent = Agent(
    instructions="You are helpful",
    context=ManagerConfig(
        tool_limits={
            "tavily_search": 2000,    # Search: 2K chars
            "tavily_extract": 5000,   # Full page: 5K chars
            "code_executor": 10000,   # Code output: 10K chars
        },
        protected_tools=["file_read"],  # Never pruned
    )
)

Session Deduplication

Prevents duplicate content across agents in multi-agent workflows:

Diagram

Multi-Agent Policies

Control how context is shared between agents:

Diagram

Mode	Description	Use Case
`NONE`	No context shared	Independent agents
`SUMMARY`	Summarized context	Reduce tokens
`FULL`	Full context (bounded)	Continuity needed

python

from praisonaiagents import ContextPolicy, ContextShareMode

policy = ContextPolicy(
    share=True,
    share_mode=ContextShareMode.SUMMARY,
    max_tokens=5000,
    preserve_recent_turns=3,
)

Context Monitoring

Real-time snapshots for debugging:

Diagram

python

from praisonaiagents import Agent, ManagerConfig

agent = Agent(
    instructions="You are helpful",
    context=ManagerConfig(
        monitor_enabled=True,
        monitor_path="./context.txt",
        monitor_format="human",  # or "json"
        redact_sensitive=True,
    )
)

Snapshot Output

================================================================================
PRAISONAI CONTEXT SNAPSHOT
================================================================================
Timestamp: 2026-01-24T06:00:00Z
Model: gpt-4o-mini
Model Limit: 128,000 tokens
Usable Budget: 111,616 tokens

--------------------------------------------------------------------------------
TOKEN LEDGER
--------------------------------------------------------------------------------
Segment              |     Tokens |     Budget |    Usage
--------------------------------------------------------------------------------
System Prompt        |        150 |      2,000 |    7.5%
History              |      5,230 |     84,616 |    6.2%
Tool Outputs         |      1,200 |     20,000 |    6.0%
--------------------------------------------------------------------------------
TOTAL                |      6,580 |    111,616 |    5.9%

Token Estimation

Fast offline token counting (no API calls):

python

from praisonaiagents import (
    estimate_tokens_heuristic,
    estimate_messages_tokens,
)

# Estimate text tokens
tokens = estimate_tokens_heuristic("Hello world!")  # ~3

# Estimate message tokens
messages = [
    {"role": "user", "content": "Hello"},
    {"role": "assistant", "content": "Hi!"},
]
tokens = estimate_messages_tokens(messages)  # ~12

Content Type	Accuracy
English text	~90-95%
Code	~85-90%
Non-ASCII	~80-85%

Fast Context (Code Search)

Rapid parallel code search for AI agents:

Diagram

python

from praisonaiagents.context.fast import FastContext

fc = FastContext(
    workspace_path=".",
    max_turns=4,
    max_parallel=8,
)

result = fc.search("find authentication handlers")
print(f"Found {result.total_files} files in {result.search_time_ms}ms")

Feature	Value
Search Latency	100-200ms
Cache Hit	< 1ms
Parallel Speedup	2-5x

CLI Commands

bash

# Enable context in chat
praisonai chat --context

# Set strategy
praisonai chat --context-strategy smart

# Set threshold
praisonai chat --context-threshold 0.8

# Enable monitoring
praisonai chat --context-monitor

In-Session Commands

Command	Description
`/context on`	Enable monitoring
`/context off`	Disable monitoring
`/context stats`	Show token ledger
`/context dump`	Write snapshot now
`/context compact`	Force optimization

Configuration Reference

ManagerConfig Options

Option	Type	Default	Description
`auto_compact`	bool	`True`	Auto-optimize on threshold
`compact_threshold`	float	`0.8`	Trigger at this usage %
`strategy`	str	`"smart"`	Optimization strategy
`output_reserve`	int	Model-specific	Reserved for output
`llm_summarize`	bool	`False`	Use LLM for summarization
`tool_limits`	dict	`{}`	Per-tool token limits
`protected_tools`	list	`[]`	Tools never pruned
`monitor_enabled`	bool	`False`	Enable snapshots
`redact_sensitive`	bool	`True`	Redact secrets

Environment Variables

bash

PRAISONAI_CONTEXT_OUTPUT_RESERVE=8000
PRAISONAI_CONTEXT_THRESHOLD=0.8
PRAISONAI_CONTEXT_MONITOR=true

Best Practices

Enable for tool-heavy agents

Always enable context=True for agents with tools to prevent token overflow from large search results.

Use smart strategy in production

The smart strategy combines all optimization techniques intelligently.

Set per-tool limits

Configure lower limits for verbose tools (search, web scraping) and higher limits for code execution.

Monitor during development

Enable monitor_enabled=True to debug context issues and understand token usage.

Use session deduplication

In multi-agent workflows, deduplication prevents the same content from being processed multiple times.

Memory

Persistent storage across sessions

Knowledge

Pre-loaded reference documents

Context vs Memory

When to use each system

Context vs Knowledge

Runtime vs pre-loaded data

AI Agents with Context