Skip to main content

Frequently Asked Questions

Install the correct package - autogen-agentchat

The name of Autogen package at PyPI is autogen-agentchat:

pip install autogen-agentchat~=0.2

Typical errors that you might face when using the wrong package are AttributeError: module 'autogen' has no attribute 'Agent', AttributeError: module 'autogen' has no attribute 'config_list_from_json' etc.

Set your API endpoints

This documentation has been moved here.

Use the constructed configuration list in agents

This documentation has been moved here.

How does an agent decide which model to pick out of the list?

This documentation has been moved here.

Unexpected keyword argument 'base_url'

In version >=1, OpenAI renamed their api_base parameter to base_url. So for older versions, use api_base but for newer versions use base_url.

Can I use non-OpenAI models?

Yes. You currently have two options:

Handling API Rate Limits

Setting the API Rate Limit

You can set the api_rate_limit in a config_list for an agent, which will be used to control the rate at which API requests are sent.

  • api_rate_limit (float): the maximum number of API requests allowed per second.

Handle Rate Limit Error and Timeout Error

You can set max_retries to handle rate limit error. And you can set timeout to handle timeout error. They can all be specified in llm_config for an agent, which will be used in the OpenAI client for LLM inference. They can be set differently for different clients if they are set in the config_list.

  • max_retries (int): the total number of times allowed for retrying failed requests for a single client.
  • timeout (int): the timeout (in seconds) for a single client.

Please refer to the documentation for more info.

How to continue a finished conversation

When you call initiate_chat the conversation restarts by default. You can use send or initiate_chat(clear_history=False) to continue the conversation.

max_consecutive_auto_reply vs max_turn vs max_round

  • max_consecutive_auto_reply the maximum number of consecutive auto replie (a reply from an agent without human input is considered an auto reply). It plays a role when human_input_mode is not "ALWAYS".
  • max_turns in ConversableAgent.initiate_chat limits the number of conversation turns between two conversable agents (without differentiating auto-reply and reply/input from human)
  • max_round in GroupChat specifies the maximum number of rounds in a group chat session.

How do we decide what LLM is used for each agent? How many agents can be used? How do we decide how many agents in the group?

Each agent can be customized. You can use LLMs, tools, or humans behind each agent. If you use an LLM for an agent, use the one best suited for its role. There is no limit of the number of agents, but start from a small number like 2, 3. The more capable is the LLM and the fewer roles you need, the fewer agents you need.

The default user proxy agent doesn't use LLM. If you'd like to use an LLM in UserProxyAgent, the use case could be to simulate user's behavior.

The default assistant agent is instructed to use both coding and language skills. It doesn't have to do coding, depending on the tasks. And you can customize the system message. So if you want to use it for coding, use a model that's good at coding.

Why is code not saved as file?

If you are using a custom system message for the coding agent, please include something like: If you want the user to save the code in a file before executing it, put # filename: <filename> inside the code block as the first line. in the system message. This line is in the default system message of the AssistantAgent.

If the # filename doesn't appear in the suggested code still, consider adding explicit instructions such as "save the code to disk" in the initial user message in initiate_chat. The AssistantAgent doesn't save all the code by default, because there are cases in which one would just like to finish a task without saving the code.

Legacy code executor

note

The new code executors offers more choices of execution backend. Read more about code executors.

The legacy code executor is used by specifying the code_execution_config in the agent's constructor.

from autogen import UserProxyAgent

user_proxy = UserProxyAgent(
name="user_proxy",
code_execution_config={"work_dir":"_output", "use_docker":"python:3"},
)

In this example, the code_execution_config specifies that the code will be executed in a docker container with the image python:3. By default, the image name is python:3-slim if not specified. The work_dir specifies the directory where the code will be executed. If you have problems with agents running pip install or get errors similar to Error while fetching server API version: ('Connection aborted.', FileNotFoundError(2, 'No such file or directory'), you can choose 'python:3' as image as shown in the code example above and that should solve the problem.

By default it runs code in a docker container. If you want to run code locally (not recommended) then use_docker can be set to False in code_execution_config for each code-execution agent, or set AUTOGEN_USE_DOCKER to False as an environment variable.

You can also develop your AutoGen application in a docker container. For example, when developing in GitHub codespace, AutoGen runs in a docker container. If you are not developing in GitHub Codespaces, follow instructions here to install and run AutoGen in docker.

Agents keep thanking each other when using gpt-3.5-turbo

When using gpt-3.5-turbo you may often encounter agents going into a "gratitude loop", meaning when they complete a task they will begin congratulating and thanking each other in a continuous loop. This is a limitation in the performance of gpt-3.5-turbo, in contrast to gpt-4 which has no problem remembering instructions. This can hinder the experimentation experience when trying to test out your own use case with cheaper models.

A workaround is to add an additional termination notice to the prompt. This acts a "little nudge" for the LLM to remember that they need to terminate the conversation when their task is complete. You can do this by appending a string such as the following to your user input string:

prompt = "Some user query"

termination_notice = (
'\n\nDo not show appreciation in your responses, say only what is necessary. '
'if "Thank you" or "You\'re welcome" are said in the conversation, then say TERMINATE '
'to indicate the conversation is finished and this is your last message.'
)

prompt += termination_notice

Note: This workaround gets the job done around 90% of the time, but there are occurrences where the LLM still forgets to terminate the conversation.

ChromaDB fails in codespaces because of old version of sqlite3

(from issue #251)

Code examples that use chromadb (like retrieval) fail in codespaces due to a sqlite3 requirement.

>>> import chromadb
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/vscode/.local/lib/python3.10/site-packages/chromadb/__init__.py", line 69, in <module>
raise RuntimeError(
RuntimeError: Your system has an unsupported version of sqlite3. Chroma requires sqlite3 >= 3.35.0.
Please visit https://docs.trychroma.com/troubleshooting#sqlite to learn how to upgrade.

Workaround:

  1. pip install pysqlite3-binary
  2. mkdir /home/vscode/.local/lib/python3.10/site-packages/google/colab

Explanation: Per this gist, linked from the official chromadb docs, adding this folder triggers chromadb to use pysqlite3 instead of the default.

How to register a reply function

(from issue #478)

See here https://microsoft.github.io/autogen/docs/reference/agentchat/conversable_agent/#register_reply

For example, you can register a reply function that gets called when generate_reply is called for an agent.

def print_messages(recipient, messages, sender, config):
if "callback" in config and config["callback"] is not None:
callback = config["callback"]
callback(sender, recipient, messages[-1])
print(f"Messages sent to: {recipient.name} | num messages: {len(messages)}")
return False, None # required to ensure the agent communication flow continues

user_proxy.register_reply(
[autogen.Agent, None],
reply_func=print_messages,
config={"callback": None},
)

assistant.register_reply(
[autogen.Agent, None],
reply_func=print_messages,
config={"callback": None},
)

In the above, we register a print_messages function that is called each time the agent's generate_reply is triggered after receiving a message.

How to get last message ?

Refer to https://microsoft.github.io/autogen/docs/reference/agentchat/conversable_agent/#last_message

How to get each agent message ?

Please refer to https://microsoft.github.io/autogen/docs/reference/agentchat/conversable_agent#chat_messages

When using autogen docker, is it always necessary to reinstall modules?

The "use_docker" arg in an agent's code_execution_config will be set to the name of the image containing the change after execution, when the conversation finishes. You can save that image name. For a new conversation, you can set "use_docker" to the saved name of the image to start execution there.

Database locked error

When using VMs such as Azure Machine Learning compute instances, you may encounter a "database locked error". This is because the LLM cache is trying to write to a location that the application does not have access to.

You can set the cache_path_root to a location where the application has access. For example,

from autogen import Cache

with Cache.disk(cache_path_root="/tmp/.cache") as cache:
agent_a.initate_chat(agent_b, ..., cache=cache)

You can also use Redis cache instead of disk cache. For example,

from autogen import Cache

with Cache.redis(redis_url=...) as cache:
agent_a.initate_chat(agent_b, ..., cache=cache)

You can also disable the cache. See here for details.

Agents are throwing due to docker not running, how can I resolve this?

If running AutoGen locally the default for agents who execute code is for them to try and perform code execution within a docker container. If docker is not running, this will cause the agent to throw an error. To resolve this you have some options.

If you want to disable code execution entirely

  • Set code_execution_config to False for each code-execution agent. E.g.:
user_proxy = autogen.UserProxyAgent(
name="agent",
llm_config=llm_config,
code_execution_config=False)

If you want to run code execution in docker

  • Recommended: Make sure docker is up and running.

If you want to run code execution locally

  • use_docker can be set to False in code_execution_config for each code-execution agent.
  • To set it for all code-execution agents at once: set AUTOGEN_USE_DOCKER to False as an environment variable.

E.g.:

user_proxy = autogen.UserProxyAgent(
name="agent", llm_config=llm_config,
code_execution_config={"work_dir":"coding", "use_docker":False})

What should I do if I get the error "TypeError: Assistants.create() got an unexpected keyword argument 'file_ids'"?

This error typically occurs when using Autogen version earlier than 0.2.27 in combination with OpenAI library version 1.21 or later. The issue arises because the older version of Autogen does not support the file_ids parameter used by newer versions of the OpenAI API. To resolve this issue, you need to upgrade your Autogen library to version 0.2.27 or higher that ensures compatibility between Autogen and the OpenAI library.

pip install --upgrade autogen

None of the devcontainers are building due to "Hash sum mismatch", what should I do?

This is an intermittent issue that appears to be caused by some combination of mirror and proxy issues. If it arises, try to replace the apt-get update step with the following:

RUN echo "Acquire::http::Pipeline-Depth 0;" > /etc/apt/apt.conf.d/99custom && \
echo "Acquire::http::No-Cache true;" >> /etc/apt/apt.conf.d/99custom && \
echo "Acquire::BrokenProxy true;" >> /etc/apt/apt.conf.d/99custom

RUN apt-get clean && \
rm -r /var/lib/apt/lists/* && \
apt-get update -o Acquire::CompressionTypes::Order::=gz && \
apt-get -y update && \
apt-get install sudo git npm # and whatever packages need to be installed in this specific version of the devcontainer

This is a combination of StackOverflow suggestions here and here.