Multi-agent system for AI evaluations in AISI's inspect-ai framework
We provide a set of tools which allow agents to delegate tasks to sub-agents. Specifically, the init_sub_agents function return three tools:
sub_agent_specs: Return info about the sub-agent(s)run_sub_agent: Give string input to one sub-agent and execute it.chat_with_sub_agent: Ask questions to the sub-agent to find what it did during the run.
pip install multiagent-inspect
pip install openai
Instead of openai, you can use any other model provider. See https://inspect.ai-safety-institute.org.uk/models.html.
from inspect_ai.solver import basic_agent
from multiagent_inspect import SubAgentConfig, init_sub_agents
from my_inspect_tools import tool1, tool2, tool3, tool4
sub_agent_1 = SubAgentConfig(tools=[tool1, tool2], max_steps=5)
sub_agent_2 = SubAgentConfig(tools=[tool3], model="openai/gpt-4o")
main_agent=basic_agent(
init=init_sub_agents([sub_agent_1, sub_agent_2]),
tools=[tool4],
)