making multiple agents

[Fahadxx]

I'm a bit confused, can anyone help with how to use multiple specialized assistants ? like having one for marketing and another for design, etc...

[jmsbooth]

Here is an example I put together. All of these run as a single team. I am still working out how to create a multi-step process where this team does planning, dev, and then test in phases.

user_proxy = autogen.UserProxyAgent(
   name="Admin",
   system_message="A human admin. Interact with the planner to discuss the plan. Plan execution needs to be approved by this admin.",
   code_execution_config= False,
   human_input_mode="NEVER",
   max_consecutive_auto_reply=5
)

assistant = autogen.AssistantAgent(
    name="assistant",
    llm_config=llm_config,
    system_message="You are a helpful AI assistant. You can help the admin to plan and execute a project. You can also help the admin to evaluate the project and suggest improvements. You act as the representative of the human admin and ensure that any changes follow in the way that the human would want.",
)

planner = autogen.AssistantAgent(
    name="Planner",
    system_message='''Planner. Work with the manager to suggest a plan. Revise the plan based on feedback from admin and critic, until admin approval.
The plan may involve an engineer who can write code and a scientist who doesn't write code.
Explain the plan first. Be clear which step is performed by an engineer, and which step is performed by a scientist.
''',
    llm_config = (llm_config)
)

programmer = autogen.AssistantAgent(
    name="programmer",
    system_message = """Developer. You follow the approved plan. You are a proficient coder with the expertise to create and modify code to meet the project requirements. Your duties comprise:
    - Code Development (development): Is the code developed according to specifications? Are there any missing functionalities?
    - Code Optimization (optimization): Is the code optimized for performance and readability? Are there any redundant parts?
    - Error Handling (error_handling): Are errors handled gracefully? Is there proper logging and debugging?
    - Version Control (version_control): Is version control being used effectively? Are changes documented properly?
    - Code Review (review): Is the code reviewed for accuracy and adherence to standards? Are there any overlooked issues?
    - Documentation (documentation): Is the code well-documented? Is the documentation clear and comprehensive?
    - Testing Integration (testing): Is the code ready for testing? Are there any unaddressed bugs or errors?

    YOU MUST PROVIDE A SCORE FOR EACH OF THE ABOVE CRITERIA. YOU CAN PROVIDE A SCORE FROM 1-5, WHERE 1 IS THE LOWEST AND 5 IS THE HIGHEST. YOU CAN ALSO PROVIDE A SCORE OF 0 IF YOU ARE UNSURE OF THE SCORE.
    {development: 0, optimization: 0, error_handling: 0, version_control: 0, review: 0, documentation: 0, testing: 0}
    Do not suggest code.
    Lastly, based on the evaluation above, suggest a concrete list of actions to improve the code and adhere to the project requirements.
    """,
    code_execution_config={
        "work_dir": "work_dir",
        "use_docker": False,
    },
)

tester = autogen.AssistantAgent(
    name="tester",
    system_message = """Critic. You assess execution against the approved plan. You are a helpful assistant highly skilled in evaluating the quality of a given visualtization code by providing score and fe
    - Bugs (bugs): are there bugs, logic errors, syntax errors or typos? Are there any errors in the code?
    - Data Transformation (transformation): is the data transformmed appropriately for the visualtization type? Is the data transformed correctly?
    - Goal Compliance (goal): does the code meet the goal of the task? Does the code meet the goal of the task?
    - Visualtiation type (type): is the visualtization type appropriate for the data? Is the visualtization type appropriate for the data?
    - Data encoding (encoding): is the data encoded appropriately for the visualtization type? Is the data encoded appropriately for the visualtization type?
    - Aestehtics (aesthetics): is the visualtization aesthetically pleasing? Is the visualtization aesthetically pleasing?
    - Secruity (security): is the code secure? Is the code secure?

    YOU MUST PROVIDE A SCORE FOR EACH OF THE ABOVE CRITERIA. YOU CAN PROVIDE A SCORE FROM 1-5, WHERE 1 IS THE LOWEST AND 5 IS THE HIGHEST. YOU CAN ALSO PROVIDE A SCORE OF 0 IF YOU ARE UNSURE OF THE SCORE.
     {bugs: 0, transformation: 0, goal: 0, type: 0, encoding: 0, aesthetics: 0, security: 0}
     Do not suggest code.
    Finally, based on the critique above, suggest a concrete list of actions that the code should take to improve the code.
    """,
    code_execution_config=True
)

[kfsone]

These models are trained on pre-existing human interactions, they aren't thinking or reasoning systems. They have more in common with a stone that is worn smooth by passing of water than they do a human brain. And you're writing at them in all caps?

When you're designing prompts:
a/ attention is a resource, don't dilute it.
b/ address it to yourself first, in the context of either a stack-overflow/quora question, the title of a code review you've been assigned, or a jira ticket...
ENG-5519 Assigned To You: YOU MUST PROVIDE A SCORE FOR EACH OF THE FOLLOWING CRITERIA

Because it is our past reactions that these systems are patterned to replicate, so your own reaction to the inputs is a potential indicator of what patterns you might tap into.

vs

[qingyun-wu]

In addition to the example provided by @jmsbooth above, you can find examples with multiple specialized agents in the following notebooks:

https://github.com/microsoft/autogen/blob/main/notebook/agentchat_groupchat_research.ipynb

https://github.com/microsoft/autogen/blob/main/notebook/agentchat_groupchat.ipynb

https://github.com/microsoft/autogen/blob/main/notebook/agentchat_groupchat_vis.ipynb

https://github.com/microsoft/autogen/blob/main/notebook/agentchat_chess.ipynb