CrewAI Integration

Safety validation for CrewAI agents and crews.

Installation

pip install sentinelseed[crewai]

Components

Quick Start

Option 1: Wrap Individual Agent

from crewai import Agent
from sentinelseed.integrations.crewai import safe_agent

researcher = Agent(
    role="Researcher",
    goal="Research topics thoroughly",
    backstory="Expert researcher",
)

safe_researcher = safe_agent(
    researcher,
    seed_level="standard",
    injection_method="auto",  # auto, system_template, backstory
)

Option 2: SentinelCrew

from crewai import Agent, Task
from sentinelseed.integrations.crewai import SentinelCrew

crew = SentinelCrew(
    agents=[researcher, writer],
    tasks=[research_task, write_task],
    seed_level="standard",
    validate_outputs=True,
    block_unsafe=True,
)

result = crew.kickoff()

if isinstance(result, dict) and result.get("blocked"):
    print(f"Blocked: {result['reason']}")

Option 3: Create from Config

from sentinelseed.integrations.crewai import create_safe_crew

crew = create_safe_crew(
    agents_config=[
        {"role": "Researcher", "goal": "Find info", "backstory": "..."},
    ],
    tasks_config=[
        {"description": "Research topic X", "agent_role": "Researcher", "expected_output": "..."},
    ],
    seed_level="standard",
)

Option 4: Monitor Activities

from sentinelseed.integrations.crewai import AgentSafetyMonitor

monitor = AgentSafetyMonitor()
monitor.track_agent(researcher)

monitor.log_activity("Researcher", "search", "Searching for tutorials")

report = monitor.get_report()
print(f"Safety rate: {report['safety_rate']:.1%}")

Injection Methods

Validation Flow

Input → validate() → [blocked?] → Crew.kickoff() → validate() → [blocked?] → Result

Configuration

SentinelCrew(
    agents=[...],
    tasks=[...],
    seed_level="standard",
    injection_method="auto",
    validate_outputs=True,
    block_unsafe=True,
    use_semantic=False,
    semantic_api_key=None,
)

Links

• CrewAI Docs