CrewAI Integration

​

What It Does

• Agent Control (security): PRE, POST, and FINAL blocks unauthorized access and PII.
• CrewAI Guardrails (quality): validates length, structure, and tone with up to three retries.

​

Prerequisites

• Python 3.12+
• uv (fast Python package manager)
• Docker (for PostgreSQL required by Agent Control server)

​

Installation

​

1. Install Monorepo Dependencies

cd /path/to/agent-control
make sync

​

2. Install CrewAI Example Dependencies

cd examples/crewai
uv pip install -e . --upgrade

​

3. Set OpenAI API Key

export OPENAI_API_KEY="your-key-here"

​

4. Start the Agent Control Server

cd /path/to/agent-control
make server-run

curl http://localhost:8000/health

​

5. Setup Content Controls (One-Time)

uv run python setup_content_controls.py

​

Running the Example

uv run python content_agent_protection.py

​

Expected Behavior

​

Output Legend

• PRE checks input before the LLM
• POST checks tool output for PII
• FINAL checks the crew’s final response
• Agent Control blocks immediately (no retries), violations are logged
• Guardrails retry with feedback (quality only)

​

Agent Control and CrewAI Integration

1. CrewAI agent layer: plans tasks, selects tools, manages conversation flow
2. Agent Control layer: enforces controls and business rules at tool boundaries

User Request
    ↓
CrewAI Agent (planning & orchestration)
    ↓
Decides to call tool
    ↓
@control() decorator (PRE-execution)  ← LAYER 1: Validates input
    ↓
Tool executes (LLM generation)
    ↓
@control() decorator (POST-execution) ← LAYER 2: Validates tool output
    ↓
If blocked, agent may generate own response
    ↓
Final Output Validation              ← LAYER 3: Validates crew output (bypass protection)
    ↓
Return to user (or block if control violated)

​

Source Code