Skip to main content
Automatic safety controls for AWS Strands agents using hooks - no decorators needed.

​
Quick Start

Prerequisites: Python 3.12+, OpenAI API key, AgentControl server 
# 1. Install
cd examples/strands_agents
uv pip install -e .

# 2. Configure
cp .env.example .env
# Edit .env: Add OPENAI_API_KEY and AGENT_CONTROL_URL

# 3. Start server (Terminal 1)
curl -fsSL https://raw.githubusercontent.com/agentcontrol/agent-control/docker-compose.yml | docker compose -f - up -d

# 4. Setup controls (Terminal 2)
cd examples/strands_agents/interactive_demo
uv run setup_interactive_controls.py

# 5. Run demo (Terminal 3)
streamlit run interactive_support_demo.py
Open http://localhost:8501 and click test buttons to see safety controls in action.

​
Available Demos

​
Interactive Demo

Customer support agent with PII blocking and SQL injection prevention. Shows real-time safety checks in Streamlit UI.

​
Steering Demo

Banking email agent with PII redaction. Combines AgentControl Hook (deny on tool calls) and Strands Steering (steer on LLM draft) for layered governance. Uses a two-phase draft to send flow so steer can guide before tool calls.

​
How It Works

AgentControlHook provides automatic safety without code changes. 
from agent_control.integrations.strands import AgentControlHook
from strands import Agent

import agent_control
agent_control.init(agent_name="my-customer-agent")

hook = AgentControlHook(agent_name="my-customer-agent")

agent = Agent(
    model=model,
    system_prompt="...",
    tools=[...],
    hooks=[hook]
)
Hook intercepts events, the server evaluates controls, and the hook enforces the decision. For steer actions, the steering handler converts AgentControl steer into a Strands Guide() retry.

​
Controls

​
LLM Controls

Apply to all model interactions: 
{
    "scope": {"step_types": ["llm"], "stages": ["pre"]},
    "condition": {
        "selector": {"path": "input"},
        "evaluator": {"name": "regex", "config": {"pattern": r"\d{3}-\d{2}-\d{4}"}},
    },
    "action": {"decision": "deny"}
}

​
Tool Controls

Target specific tools: 
{
    "scope": {
        "step_types": ["tool"],
        "step_names": ["lookup_order"],
        "stages": ["pre"]
    },
    "condition": {
        "selector": {"path": "input"},
        "evaluator": {"name": "regex", "config": {"pattern": r"ORD-\d+"}},
    },
    "action": {"decision": "deny"}
}
Hook automatically extracts tool names from events - no decorators needed.

​
Architecture

User Input
  ↓
Strands fires event (BeforeToolCallEvent, AfterModelCallEvent, etc.)
  ↓
AgentControlHook intercepts → Creates Step → Calls AgentControl server
  ↓
Server evaluates controls → Returns safe or unsafe
  ↓
Hook enforces decision: Continue or Block

​
Integration Patterns

​
Basic Setup

hook = AgentControlHook(
    agent_name="my-customer-agent",
    enable_logging=True
)

​
Steering Integration

  • AgentControlHook for tool-stage deny (hard blocks)
  • AgentControlSteeringHandler for LLM steer to Guide() (corrective guidance)
See steering_demo/README.md for a complete implementation.

​
Source Code

View the complete example with all scripts and setup instructions: Strands Agents Integration Examples

​
Troubleshooting

​
AgentControl not initialized

Run agent_control.init() before creating the hook.

​
Controls not triggering

  • Server running? curl http://localhost:8000/health
  • Controls exist? Re-run setup_*_controls.py

​
Import errors

Install dependencies: uv sync or pip install -e .

​
Files

  • interactive_demo/ - Customer support demo with PII/SQL injection blocking
  • steering_demo/ - Banking email demo with PII redaction