agentintent

AgentIntent

AI Should Do What It Says. We Make Sure It Does.

AgentIntent is the trust infrastructure for AI agents. Register your agents, declare their intent, and monitor their behavior to build trust.

Installation

pip install agentintent

With LangChain support:

pip install agentintent[langchain]

Quick Start

LangChain (2 lines of code)

from langchain.agents import AgentExecutor
from agentintent import AgentIntentCallback

# Just add our callback - that's it!
agent = AgentExecutor(
    agent=my_agent,
    tools=my_tools,
    callbacks=[AgentIntentCallback("ai_xxxxx")]  # ← Add this line
)

# Use your agent as normal
result = agent.invoke({"input": "Do something"})

With Intent Manifest (recommended)

Define what your agent should do, and we'll make sure it does:

from agentintent import AgentIntent, IntentManifest

# Define what your agent is supposed to do
manifest = IntentManifest(
    name="ExpenseBot",
    description="Processes expense reports",
    
    # Tools this agent is allowed to use
    allowed_tools=["email_read", "ocr", "form_submit"],
    
    # Tools this agent must NEVER use
    forbidden_tools=["code_execute", "file_delete", "send_email"],
    
    # Block violations instead of just warning
    enforcement_mode="block"
)

# Initialize AgentIntent
ai = AgentIntent(api_key="ai_xxxxx", manifest=manifest)

# Get the callback for LangChain
callback = ai.langchain_callback()

# Use with your agent
agent = AgentExecutor(
    agent=my_agent,
    tools=my_tools,
    callbacks=[callback]
)

# If agent tries to use a forbidden tool, it will be blocked
result = agent.invoke({"input": "Process this expense report"})

Universal Wrapper (any framework)

from agentintent import monitor

# Wrap any agent or function
monitored_agent = monitor(your_agent, api_key="ai_xxxxx")

# Or use as a decorator
@monitor(api_key="ai_xxxxx")
def my_agent_function(input):
    # Your agent logic
    return result

Features

🛡️ Intent Declaration

Tell us what your agent should do:

manifest = IntentManifest(
    name="CustomerServiceBot",
    allowed_tools=["search_kb", "send_response"],
    forbidden_tools=["access_billing", "modify_account"],
    allowed_api_patterns=["api.company.com/*"],
    forbidden_api_patterns=["*.stripe.com/*"],
)

📊 Automatic Monitoring

Every action is tracked:

• LLM calls (model, tokens, duration)
• Tool usage (which tools, inputs/outputs)
• API calls (endpoints, patterns)
• Agent decisions (reasoning, tool selection)

⚡ Real-Time Enforcement

Block bad behavior before it happens:

# enforcement_mode options:
# - "block": Raise exception on violation
# - "warn": Log warning but allow
# - "log": Just log, no warnings

manifest = IntentManifest(
    name="SecureBot",
    forbidden_tools=["dangerous_tool"],
    enforcement_mode="block"  # Will raise EnforcementError
)

🏆 Trust Score

Build reputation over time:

# Get your agent's trust score
score = ai.get_trust_score()
# {
#     "score": 94,
#     "tier": "TRUSTED",
#     "breakdown": {
#         "intent_alignment": 98,
#         "behavioral_consistency": 96,
#         "developer_reputation": 91,
#         "security_posture": 89
#     }
# }

Supported Frameworks

Performance

AgentIntent is designed for zero-impact monitoring:

• < 1ms overhead per action (local enforcement)
• Async logging (never blocks your agent)
• Batched uploads (efficient network usage)
• Fail-safe (if our cloud is down, your agent keeps running)

Configuration

Environment Variables

export AGENTINTENT_API_KEY="ai_xxxxx"
export AGENTINTENT_BASE_URL="https://api.agentintent.ai/v1"  # optional

Manifest from File

# manifest.yaml
manifest = IntentManifest.from_file("manifest.yaml")

# manifest.yaml
name: MyAgent
version: "1.0.0"
description: Does useful things

allowed_tools:
  - search
  - calculator

forbidden_tools:
  - code_execute
  - file_delete

limits:
  max_llm_calls_per_minute: 60
  max_tool_calls_per_minute: 120

enforcement_mode: warn

API Reference

AgentIntent

ai = AgentIntent(
    api_key="ai_xxxxx",
    manifest=manifest,           # Optional: IntentManifest
    enforcement_mode="warn",     # Optional: override manifest
    cloud_enabled=True,          # Optional: disable cloud logging
)

# Get callbacks/wrappers
callback = ai.langchain_callback()

# Manual logging
ai.log_action("tool_call", {"tool": "search", "query": "..."})

# Check enforcement
result = ai.check_tool("dangerous_tool")
if result.blocked:
    print(f"Blocked: {result.reason}")

# Get metrics
metrics = ai.get_metrics()

# Cleanup
ai.close()

IntentManifest

manifest = IntentManifest(
    name="MyAgent",
    version="1.0.0",
    description="What this agent does",
    
    # Tool permissions
    allowed_tools=["tool1", "tool2"],
    forbidden_tools=["bad_tool"],
    
    # API permissions (regex patterns)
    allowed_api_patterns=["api.safe.com/*"],
    forbidden_api_patterns=["*.dangerous.com/*"],
    
    # Data permissions
    allowed_data_sources=["public_db"],
    forbidden_data_sources=["private_db", "credentials"],
    
    # Rate limits
    limits={
        "max_llm_calls_per_minute": 60,
        "max_tool_calls_per_minute": 120,
    },
    
    # Enforcement
    enforcement_mode="block",  # or "warn" or "log"
)

License

MIT License - see the LICENSE file in the repository.

Links

• Website: https://agentintent.ai
• Documentation: https://docs.agentintent.ai
• GitHub: https://github.com/agentintent/agentintent-python
• Discord: https://discord.gg/agentintent