LangChain SQL Agent with Agent Control

​

Setup

​

1. Start the Agent Control Server (for remote execution)

# Kill any old servers
pkill -f "uvicorn agent_control_server"

# Start the server
make server-run
# OR: uv run --package agent-control-server uvicorn agent_control_server.main:app --port 8000

curl http://localhost:8000/api/v1/evaluators | python -m json.tool
# Should show: {"sql": {"name": "sql", "version": "1.0.0", ...}, ...}

​

2. Set OpenAI API Key

export OPENAI_API_KEY="your-key-here"

​

3. Install pre-reqs

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

​

4. Setup SQL Controls (One-Time)

uv run setup_sql_controls.py

• SQL safety control (blocks DROP, DELETE, TRUNCATE, ALTER, GRANT)
• Policy with the control
• Assigns policy to the SQL agent

For local execution, create the control with execution: "sdk" in setup_sql_controls.py (see sql_control_data_sdk) and enable AGENT_CONTROL_LOCAL_EVAL=true when running the agent.

​

Running the Example

uv run sql_agent_protection.py

​

Local vs Remote Control Execution

• Default mode
• Requires running the Agent Control server
• Uses @control() to call /api/v1/evaluation on the server

• Set AGENT_CONTROL_LOCAL_EVAL=true
• Controls must be configured with execution: "sdk"
• Uses agent_control.check_evaluation_with_local(...) before executing the tool

export AGENT_CONTROL_LOCAL_EVAL=true
uv run sql_agent_protection.py

​

Expected Behavior

✅ Safety check passed for: SELECT * FROM Track LIMIT 3
[Query executes successfully]

✗ Execution blocked!
Error: Control evaluation failed...
[Query is blocked, table is NOT dropped]

​

How It Works

​

1. The @control() Decorator

async def _execute_query(query: str):
    return query_tool.invoke(query)

# Set tool name (required for tool step detection)
_execute_query.name = "sql_db_query"
_execute_query.tool_name = "sql_db_query"

# Apply decorators
safe_query_tool = tool(
    "sql_db_query",
    description="Execute SQL query with safety checks"
)(control()(_execute_query))

​

2. Automatic Tool Call Detection

1. Detects this is a tool (via name attribute)
2. Creates a Step payload with type="tool", name, and input
3. Sends to server for evaluation before execution
4. Blocks execution if control triggers deny action

​

3. SQL Control Execution

• Parses the query
• Checks for blocked operations (DROP, DELETE, etc.)
• Validates LIMIT clauses
• Returns deny/observe decision

​

4. Fail-Safe Error Handling

• ✅ Execution is BLOCKED (fail-safe)
• ✅ RuntimeError is raised
• ❌ Query never executes

​

Troubleshooting

​

”Evaluator ‘sql’ not found”

# Kill old server
pkill -f "uvicorn agent_control_server"

# Restart server
cd server && make run

​

“Policy ‘sql-protection-policy’ already exists”

​

DROP TABLE still executes

1. Server not running or evaluators not loaded (remote mode)
2. Control not assigned to agent’s policy
3. Control data missing/invalid (control not returned to agent)
4. Local mode enabled but control is still execution: "server"

1. Restart server with make run
2. Re-run setup_sql_controls.py
3. Verify decorator code is up-to-date

​

Architecture

LangChain Agent
    ↓
@tool("sql_db_query")  ← Marks as LangChain tool
@control()            ← Applies server-side validation
    ↓
Step Payload: {
  "type": "tool",
  "name": "sql_db_query",
  "input": {"query": "DROP TABLE..."}
}
    ↓
Agent Control Server
    ↓
SQL Evaluator Execution
    ↓
DENY (blocks DROP) or PASS (safe query)
    ↓
Back to Decorator
    ↓
Raise ControlViolationError (DENY) or Execute (PASS)

​

Files

• sql_agent_protection.py - Main SQL agent with @control() decorator
• setup_sql_controls.py - One-time setup script for controls/policy
• pyproject.toml - Dependencies and configuration
• README.md - This file

​

Key Security Features

1. Fail-Safe by Default: Errors block execution
2. Server-Side Validation: Remote controls enforced centrally
3. SDK-Side Validation: Local controls run before tool execution
4. Automatic Detection: Decorator auto-detects tool vs LLM calls

​

Source Code