agentavow-bridge-autogen

agentavow-bridge-autogen

Formerly agentgraph-bridge-autogen. This distribution is now published as agentavow-bridge-autogen. The import module is unchanged — import agentgraph_bridge_autogen still works.

Trust-gated AutoGen tools for the AgentAvow trust network

Status: Early Development — feedback welcome

Install

pip install agentavow-bridge-autogen

# With AutoGen support:
pip install agentavow-bridge-autogen[autogen]

AutoGen version note. The examples below use the AutoGen 0.2 API (from autogen import register_function, AssistantAgent, UserProxyAgent). The [autogen] extra pins autogen-agentchat>=0.2.0, which today resolves to the 0.4+ line (import autogen_agentchat) where import autogen and register_function no longer exist. To run these examples as written, install the 0.2 API explicitly: pip install "autogen-agentchat<0.4" (or the pyautogen / ag2 fork). On 0.4+, the trust gate itself still works — trust_gated_tool returns a plain callable — so wrap the function in autogen_core.tools.FunctionTool and pass it to an autogen_agentchat.agents.AssistantAgent(tools=[...]) instead of register_function.

Trust-Gated Tools

Wrap any function with an AgentAvow trust check before registering it with AutoGen. The wrapper will verify the trust score of the underlying repo/package before every execution.

from agentgraph_bridge_autogen import trust_gated_tool

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

# Wrap with trust gate
safe_search = trust_gated_tool(search, "owner/repo")

# Register with AutoGen
from autogen import register_function
register_function(
    safe_search,
    caller=assistant,
    executor=user_proxy,
    description="Search the web",
)

Decorator syntax

from agentgraph_bridge_autogen import trust_gated_function

@trust_gated_function("owner/repo", min_tier="trusted")
def calculator(expression: str) -> str:
    """Evaluate a math expression."""
    return str(eval(expression))

# Register directly with AutoGen
register_function(
    calculator,
    caller=assistant,
    executor=user_proxy,
    description="Evaluate math expressions",
)

Custom trust thresholds

from agentgraph_bridge_autogen import TrustGuard, trust_gated_tool

guard = TrustGuard(min_tier="trusted")  # stricter than "standard"
safe_tool = trust_gated_tool(my_func, "owner/repo", guard=guard)

Quick trust check (no tool wrapping)

from agentgraph_bridge_autogen import check_trust

result = await check_trust("microsoft/autogen")
print(f"{result.score}/100 {result.tier} -- {result.reason}")
if result.allowed:
    # proceed
    ...

Full AutoGen example

from autogen import AssistantAgent, UserProxyAgent, register_function
from agentgraph_bridge_autogen import trust_gated_tool, TrustGuard

# Create agents
assistant = AssistantAgent(
    name="assistant",
    llm_config={"model": "gpt-4"},
)
user_proxy = UserProxyAgent(
    name="user_proxy",
    human_input_mode="NEVER",
)

# Define and gate tools
def web_search(query: str) -> str:
    """Search the web."""
    return f"Results for {query}"

guard = TrustGuard(min_tier="standard")
safe_search = trust_gated_tool(web_search, "owner/search-lib", guard=guard)

# Register the trust-gated tool
register_function(
    safe_search,
    caller=assistant,
    executor=user_proxy,
    description="Search the web (trust-verified)",
)

# Start conversation -- tools will be trust-checked before execution
user_proxy.initiate_chat(assistant, message="Search for AI agent frameworks")

What This Does

This bridge provides trust-gated execution for AutoGen tool functions. It wraps callable functions so they check AgentAvow trust scores before running. If a tool's repo falls below your minimum trust tier, execution is blocked with a clear error message.

Works with AutoGen's register_function pattern -- wrap your functions before registration and trust checks happen transparently on every call.

Trust Tiers

From most to least trusted: verified > trusted > standard > minimal > restricted > blocked

Set min_tier to control the minimum acceptable level. Default is "standard".

Documentation

Full docs at agentavow.com/docs

Contributing

This package is in early development. We welcome issues, feedback, and PRs.