Skip to content

AutoGen Integration

Wrap AutoGen tools so every function call in your multi-agent conversation passes through SafeAI's policy engine -- secret detection, PII filtering, tool contracts, and audit logging are enforced transparently.

Auto-generate AutoGen integration

Let the intelligence layer generate SafeAI wrappers for your AutoGen agents: 

safeai intelligence integrate --target autogen --path . --apply

Install

uv pip install safeai pyautogen

Quick Start

from safeai import SafeAI

ai = SafeAI()
adapter = ai.autogen_adapter()
safe_tool = adapter.wrap_tool("execute_code", code_tool, agent_id="coder")

Detailed Usage

Creating the Adapter

from safeai import SafeAI

ai = SafeAI.from_config("safeai.yaml")
adapter = ai.autogen_adapter()   # returns SafeAIAutoGenAdapter

You can also import the adapter directly:

from safeai.middleware.autogen import SafeAIAutoGenAdapter

Wrapping Tools

AutoGen registers tools as Python functions. The adapter wraps these functions so that every call is intercepted:

def search_web(query: str) -> str:
    """Search the web for information."""
    return f"Results for: {query}"

# Wrap with SafeAI
safe_search = adapter.wrap_tool(
    name="search_web",
    tool=search_web,
    agent_id="research-agent",
)

Request and response interception

The adapter intercepts both the request (function arguments) and the response (return value). Inputs are scanned before execution; outputs are guarded before they reach the agent.

Registering Wrapped Tools with AutoGen Agents

import autogen

assistant = autogen.AssistantAgent(
    name="assistant",
    llm_config={"config_list": config_list},
)

user_proxy = autogen.UserProxyAgent(
    name="user_proxy",
    human_input_mode="NEVER",
)

# Register the SafeAI-wrapped function
user_proxy.register_function(
    function_map={
        "search_web": safe_search,
    }
)

Full Example

import autogen
from safeai import SafeAI

# 1. SafeAI setup
ai = SafeAI.from_config("safeai.yaml")
adapter = ai.autogen_adapter()

# 2. Define tools
def execute_code(code: str) -> str:
    """Execute Python code in a sandbox."""
    # ... sandboxed execution ...
    return "execution result"

def read_file(path: str) -> str:
    """Read a file from disk."""
    with open(path) as f:
        return f.read()

# 3. Wrap tools
safe_execute = adapter.wrap_tool("execute_code", execute_code, agent_id="coder")
safe_read = adapter.wrap_tool("read_file", read_file, agent_id="coder")

# 4. Configure AutoGen agents
config_list = [{"model": "gpt-4", "api_key": "..."}]

assistant = autogen.AssistantAgent(
    name="assistant",
    system_message="You are a helpful coding assistant.",
    llm_config={"config_list": config_list},
)

user_proxy = autogen.UserProxyAgent(
    name="user_proxy",
    human_input_mode="NEVER",
    code_execution_config=False,
)

# 5. Register wrapped tools
user_proxy.register_function(
    function_map={
        "execute_code": safe_execute,
        "read_file": safe_read,
    }
)

# Every function call is now guarded by SafeAI
user_proxy.initiate_chat(assistant, message="Read config.yaml and summarize it")

Credential leaks are caught

If the assistant tries to read a file containing API keys and pass them to execute_code, SafeAI detects the secret in the input and blocks the call.

Multi-Agent Conversations

AutoGen excels at multi-agent group chats. SafeAI can enforce different policies per agent:

# Different agent IDs get different permissions
safe_code_tool = adapter.wrap_tool("execute_code", execute_code, agent_id="coder")
safe_review_tool = adapter.wrap_tool("execute_code", execute_code, agent_id="reviewer")

# safeai.yaml
tool_contracts:
  execute_code:
    allowed_agents: ["coder"]       # reviewer cannot execute code
    blocked_patterns:
      - "os.system"
      - "subprocess"

Configuration

# safeai.yaml
policy:
  default_action: block
  secret_detection:
    enabled: true
  pii_protection:
    enabled: true
    action: redact

tool_contracts:
  execute_code:
    allowed_agents: ["coder"]
    max_calls_per_minute: 5
  read_file:
    allowed_agents: ["coder", "reviewer"]
    blocked_patterns:
      - "/etc/shadow"
      - ".env"

audit:
  enabled: true
  log_inputs: true
  log_outputs: true

API Reference

Class Description
SafeAIAutoGenAdapter Main adapter returned by ai.autogen_adapter()
adapter.wrap_tool() Wrap a single function with policy enforcement

See API Reference - Middleware for full signatures.

Next Steps