agent_inspect.user_proxy package

Subpackages

agent_inspect.user_proxy.utils package

Submodules

agent_inspect.user_proxy.constants module

agent_inspect.user_proxy.llm_proxy_agent module

class agent_inspect.user_proxy.llm_proxy_agent.LLMProxyAgent(llm_client, config=None)[source]

Bases: ABC

Abstract class which should be extended for actual implementation of LLM agent.

Parameters:

llm_client (LLMClient) – the connection to the llm client for response generation.
config (Optional[Dict[str, Any]]) – configuration for LLM agent initialization. Default to None.

abstract async generate_message_from_chat_history(chat_history)[source]

This is an abstract method and should be implemented in a concrete class.

Parameters:: chat_history (ChatHistory) – a ChatHistory object containing the conversation history.
Return type:: UserProxyMessage
Returns:: a UserProxyMessage object containing the LLM agent response.

agent_inspect.user_proxy.templates module

agent_inspect.user_proxy.user_proxy_agent module

class agent_inspect.user_proxy.user_proxy_agent.UserProxyAgent(llm_client, task_summary, terminating_conditions, agent_description='', initial_message='', config=None)[source]

Bases: LLMProxyAgent

User proxy (a.k.a. simulated user) class which generates the user utterances during a dynamic conversation with the AI agent. The dynamic user utterances are generated based on the user task instruction and user persona (e.g., expert or non-expert) prompt templates via a two-step process—reflection followed by response generation.

Parameters:

llm_client (LLMClient) – the connection to the llm client for user utterances generation.
task_summary (str) – a user task instruction describing the summary of the task user wants the AI agent to complete.
terminating_conditions (List[TerminatingCondition]) – a List [TerminatingCondition] object where each element in the list is a terminating condition for the user proxy to exit the user-agent conversation early.
agent_description (str) – the description of the AI agent that will interact with the user proxy, provided as additional context for the user proxy. Default to empty string.
initial_message (str) – a static message that is used as the user proxy’s initial message (if available) to the AI agent. Default to empty string.
config (Optional[Dict[str, Any]]) –
Default to None. Configuration options:
- use_expert_agent: a bool flag to indicate whether the user proxy should use an expert persona; otherwise, it uses a non-expert persona. Default to True to use an expert persona.
- terminating_condition_mode: a str variable to indicate whether the user proxy is configured with single terminating condition or multiple terminating conditions. The value can be either "single" or "multiple". Default to "single".

async generate_message_from_chat_history(chat_history)[source]

Generates the next user utterance given a ChatHistory object containing the user-agent conversation history as input.

Parameters:: chat_history (ChatHistory) – a ChatHistory object containing the user-agent conversation history.
Return type:: UserProxyMessage
Returns:: a UserProxyMessage object containing the next user utterance. For the first user utterance, if the input variable chat_history is None and self.initial_message is not an empty string, the method returns the self.initial_message as the initial user utterance, otherwise it generates the user utterance based on the conversation history. All subsequent user utterances are generated based on the conversation history.

Example:

>>> from agent_inspect.user_proxy import UserProxyAgent
>>> from agent_inspect.models.user_proxy import ChatHistory TerminatingCondition
>>> from agent_inspect.user_proxy.constants import USE_EXPERT_AGENT
>>> from agent_inspect.clients.azure_openai_client import AzureOpenAIClient
>>> from uuid import uuid4
>>> import asyncio
>>>
>>> user_instruct, term_condition = load_user_instruct_term(sample_path) # Load user instruction and terminating condition
>>> client = AzureOpenAIClient(model="gpt-4.1", max_tokens=4096) # create llm client for user proxy
>>> user = UserProxyAgent(
...     llm_client=client,
...     task_summary=user_instruct,
...     terminating_conditions=[
...         TerminatingCondition(check=term_condition)
...     ],
...     config={USE_EXPERT_AGENT: True}
... )
>>> chat_history = ChatHistory(id=str(uuid4()), conversations=[]) # start from an empty conversation
>>> user_response = asyncio.run(user.generate_message_from_chat_history(chat_history))
>>> print(user_response.message_str)

Module contents

class agent_inspect.user_proxy.UserProxyAgent(llm_client, task_summary, terminating_conditions, agent_description='', initial_message='', config=None)[source]

Bases: LLMProxyAgent

User proxy (a.k.a. simulated user) class which generates the user utterances during a dynamic conversation with the AI agent. The dynamic user utterances are generated based on the user task instruction and user persona (e.g., expert or non-expert) prompt templates via a two-step process—reflection followed by response generation.

Parameters:

llm_client (LLMClient) – the connection to the llm client for user utterances generation.
task_summary (str) – a user task instruction describing the summary of the task user wants the AI agent to complete.
terminating_conditions (List[TerminatingCondition]) – a List [TerminatingCondition] object where each element in the list is a terminating condition for the user proxy to exit the user-agent conversation early.
agent_description (str) – the description of the AI agent that will interact with the user proxy, provided as additional context for the user proxy. Default to empty string.
initial_message (str) – a static message that is used as the user proxy’s initial message (if available) to the AI agent. Default to empty string.
config (Optional[Dict[str, Any]]) –
Default to None. Configuration options:
- use_expert_agent: a bool flag to indicate whether the user proxy should use an expert persona; otherwise, it uses a non-expert persona. Default to True to use an expert persona.
- terminating_condition_mode: a str variable to indicate whether the user proxy is configured with single terminating condition or multiple terminating conditions. The value can be either "single" or "multiple". Default to "single".

async generate_message_from_chat_history(chat_history)[source]

Generates the next user utterance given a ChatHistory object containing the user-agent conversation history as input.

Parameters:: chat_history (ChatHistory) – a ChatHistory object containing the user-agent conversation history.
Return type:: UserProxyMessage
Returns:: a UserProxyMessage object containing the next user utterance. For the first user utterance, if the input variable chat_history is None and self.initial_message is not an empty string, the method returns the self.initial_message as the initial user utterance, otherwise it generates the user utterance based on the conversation history. All subsequent user utterances are generated based on the conversation history.

Example:

>>> from agent_inspect.user_proxy import UserProxyAgent
>>> from agent_inspect.models.user_proxy import ChatHistory TerminatingCondition
>>> from agent_inspect.user_proxy.constants import USE_EXPERT_AGENT
>>> from agent_inspect.clients.azure_openai_client import AzureOpenAIClient
>>> from uuid import uuid4
>>> import asyncio
>>>
>>> user_instruct, term_condition = load_user_instruct_term(sample_path) # Load user instruction and terminating condition
>>> client = AzureOpenAIClient(model="gpt-4.1", max_tokens=4096) # create llm client for user proxy
>>> user = UserProxyAgent(
...     llm_client=client,
...     task_summary=user_instruct,
...     terminating_conditions=[
...         TerminatingCondition(check=term_condition)
...     ],
...     config={USE_EXPERT_AGENT: True}
... )
>>> chat_history = ChatHistory(id=str(uuid4()), conversations=[]) # start from an empty conversation
>>> user_response = asyncio.run(user.generate_message_from_chat_history(chat_history))
>>> print(user_response.message_str)