• Paper: https://web3.arxiv.org/pdf/2407.15734
• Video: https://www.youtube.com/watch?v=F3usuxs2p1Y
• Related Repositories: StrictJSON (https://github.com/tanchongmin/strictjson) [Do help star this as well!]

Happy to share that the task-based agentic framework I have been working on - TaskGen - is largely complete!

Noteable features include:

• Splitting of Tasks into subtasks for bite-sized solutions for each subtask
• Single Agent with LLM Functions
• Single Agent with External Functions
• Meta Agent with Inner Agents as Functions
• Shared Variables for multi-modality support
• Retrieval Augmented Generation (RAG) over Function space
• Memory to provide additional task-based prompts for task
• Global Context for configuring your own prompts + add persistent variables
• Async mode for Agent, Function and strict_json added
• Community Uploading and Downloading of Agent and Functions

I am quite sure that this is the best open-source agentic framework for task-based execution out there! Existing frameworks like AutoGen rely too much on conversational text which is lengthy and not targeted. TaskGen uses StrictJSON (JSON parser with type checking and more!) as the core, and agents are efficient and are able to do Chain of Thought natively using JSON keys and descriptions as a guide.

I can't wait to see what this new framework can do for you!

• JSON format helps do Chain-of-Thought prompting naturally and is less verbose than free text
• JSON format allows natural parsing of multiple output fields by agents
• StrictJSON helps to ensure all output fields are there and of the right format required for downstream processing

• Created: 17 Feb 2024 - 15 Oct 2024 by John Tan Chong Min
• Co-developer / Lead Contributor: Prince Saroj
• TaskGen Paper Research Staff: Prince Saroj, Hardik Maheshwari, Bharat Runwal, Brian Lim, Richard Cottrill
• TaskGen Mentors / Funders: Ambuj Kumar, Alankrit Chona, Mehul Motani

1. Download package via command line pip install taskgen-ai==3.3.4
2. Set up your LLM and provide any API keys if needed
3. Import the required functions from taskgen and use them!

• Default model is now gpt-4o-mini if you do not specify any LLM!
• Weaker models like ChatGPT (gpt-3.5-turbo) and Llama 3 8B are consistent only if you specify very clearly what you want the Agent to do and give examples of what you want
• gpt-4o-mini, gpt-4o, Llama 3 70B and more advanced models can perform better zero-shot without much examples
• TaskGen is compatible with ChatGPT and similar models, but for more robust use, consider using gpt-4o-mini and better models

• Create an agent by entering your agent's name and description

• Agents are task-based, so they will help generate subtasks to fulfil your main task

• Agents are made to be non-verbose, so they will just focus only on task instruction (Much more efficient compared to conversational-based agentic frameworks like AutoGen)

• Agent's interactions will be stored into subtasks_completed by default, which will serve as a memory buffer for future interactions

• Inputs for Agent:

  • agent_name: String. Name of agent, hinting at what the agent does
  • agent_description: String. Short description of what the agent does
  • max_subtasks: Int. Default: 5. The maximum number of subtasks the agent can have
  • verbose: Bool. Default: True. Whether to print out agent's intermediate thoughts
  • llm: Function. The LLM to be used by the Agent
    
    

• Agent Internal Parameters:

  • Task: String. The task the agent has been assigned to - Defaults to "No task assigned"
  • Subtasks Completed: Dict. The keys are the subtask names and the values are the result of the respective subtask
  • Is Task Completed: Bool. Whether the current Task is completed
    
    

• Task Running

  • reset(): Resets the Agent Internal Parameters and Subtasks Completed. You should do this at the start of every new task assigned to the Agent to minimise potential confusion of what has been done for this task versus previous tasks
  • run(task: str, num_subtasks: int = max_subtasks): Performs the task. Do note that agent's state will not be reset, so if you want to reset it, call reset() prior to running this. Runs the task for num_subtasks steps. If not specified, we will take the max_subtasks.
    
    

• Give User Output

  • reply_user(query: str = '', stateful: bool = True): Using all information from subtasks, give a reply about the query to the user. If query is not given, then it replies based on the current task the agent is doing. If stateful is True, saves this query and reply into subtasks_completed
    
    

• Check status of Agent:

  • status(): Lists out Agent Name, Agent Description, Available Functions (default function is to use the LLM), Task, Subtasks Completed and Is Task Completed

my_agent = Agent('Helpful assistant', 'You are a generalist agent', llm = llm)

output = my_agent.run('Give me 5 words rhyming with cool, and make a 4-sentence poem using them')

Subtask identified: Find 5 words that rhyme with 'cool'

Getting LLM to perform the following task: Find 5 words that rhyme with 'cool'

pool, rule, fool, tool, school

Subtask identified: Compose a 4-sentence poem using the words 'pool', 'rule', 'fool', 'tool', and 'school'

Getting LLM to perform the following task: Compose a 4-sentence poem using the words 'pool', 'rule', 'fool', 'tool', and 'school'

In the school, the golden rule is to never be a fool. Use your mind as a tool, and always follow the pool.

Task completed successfully!

my_agent.status()

Agent Name: Helpful assistant

Agent Description: You are a generalist agent

Available Functions: ['use_llm', 'end_task']

Task: Give me 5 words rhyming with cool, and make a 4-sentence poem using them

Subtasks Completed:

Subtask: Find 5 words that rhyme with 'cool'

pool, rule, fool, tool, school

Subtask: Compose a 4-sentence poem using the words 'pool', 'rule', 'fool', 'tool', and 'school'

In the school, the golden rule is to never be a fool. Use your mind as a tool, and always follow the pool.

Is Task Completed: True

output = my_agent.reply_user()

Here are 5 words that rhyme with "cool": pool, rule, fool, tool, school. Here is a 4-sentence poem using these words: "In the school, the golden rule is to never be a fool. Use your mind as a tool, and always follow the pool."

• First define the functions, either using class Function (see Tutorial 0), or just any Python function with input and output types defined in the signature and with a docstring
• After creating your agent, use assign_functions to assign a list of functions of class Function, or general Python functions (which will be converted to AsyncFunction)
• Function names will be automatically inferred if not specified
• Proceed to run tasks by using run()

# This is an example of an LLM-based function (see Tutorial 0)
sentence_style = Function(fn_description = 'Output a sentence with words <var1> and <var2> in the style of <var3>', 
                         output_format = {'output': 'sentence'},
                         fn_name = 'sentence_with_objects_entities_emotion',
                         llm = llm)

# This is an example of an external user-defined function (see Tutorial 0)
def binary_to_decimal(binary_number: str) -> int:
    '''Converts binary_number to integer of base 10'''
    return int(str(binary_number), 2)

# Initialise your Agent
my_agent = Agent('Helpful assistant', 'You are a generalist agent')

# Assign the functions
my_agent.assign_functions([sentence_style, binary_to_decimal])

# Run the Agent
output = my_agent.run('First convert binary string 1001 to a number, then generate me a happy sentence with that number and a ball')

Subtask identified: Convert the binary number 1001 to decimal Calling function binary_to_decimal with parameters {'x': '1001'}

{'output1': 9}

Subtask identified: Generate a happy sentence with the decimal number and a ball Calling function sentence_with_objects_entities_emotion with parameters {'obj': '9', 'entity': 'ball', 'emotion': 'happy'}

{'output': 'I am so happy with my 9 balls.'}

Task completed successfully!

• Approach 1: Automatically Run your agent using run()

• Approach 2: Manually select and use functions for your task

  • select_function(task: str): Based on the task, output the next function name and input parameters
  • use_function(function_name: str, function_params: dict, subtask: str = '', stateful: bool = True): Uses the function named function_name with function_params. stateful controls whether the output of this function will be saved to subtasks_completed under the key of subtask
    
    

• Assign/Remove Functions:

  • assign_functions(function_list: list): Assigns a list of functions to the agent
  • remove_function(function_name: str): Removes function named function_name from the list of assigned functions
    
    

• Show Functions:

  • list_functions(): Returns the list of functions of the agent
  • print_functions(): Prints the list of functions of the agent
    
    

• AsyncAgent works the same way as Agent, only much faster due to parallelisation of tasks
• It can only be assigned functions of class AsyncFunction, or general Python functions (which will be converted to AsyncFunction)
• If you define your own AsyncFunction, you should define the fn_name as well if it is not an External Function
• As a rule of thumb, just add the await keyword to any function that you run with the AsyncAgent

async def llm_async(system_prompt: str, user_prompt: str):
    ''' Here, we use OpenAI for illustration, you can change it to your own LLM '''
    # ensure your LLM imports are all within this function
    from openai import AsyncOpenAI
    
    # define your own LLM here
    client = AsyncOpenAI()
    response = await client.chat.completions.create(
        model='gpt-4o-mini',
        temperature = 0,
        messages=[
            {"role": "system", "content": system_prompt},
            {"role": "user", "content": user_prompt}
        ]
    )
    return response.choices[0].message.content

# This is an example of an LLM-based function (see Tutorial 0)
sentence_style = AsyncFunction(fn_description = 'Output a sentence with words <var1> and <var2> in the style of <var3>', 
                         output_format = {'output': 'sentence'},
                         fn_name = 'sentence_with_objects_entities_emotion', # you must define fn_name for LLM-based functions
                         llm = llm_async) # use an async LLM function

# This is an example of an external user-defined function (see Tutorial 0)
def binary_to_decimal(binary_number: str) -> int:
    '''Converts binary_number to integer of base 10'''
    return int(str(binary_number), 2)

# Initialise your Agent
my_agent = AsyncAgent('Helpful assistant', 'You are a generalist agent')

# Assign the functions
my_agent.assign_functions([sentence_style, binary_to_decimal])

# Run the Agent
output = await my_agent.run('Generate me a happy sentence with a number and a ball. The number is b1001 converted to decimal')

"Because text is not enough" - Anonymous

• shared_variables is a dictionary, that is initialised in Agent (default empty dictionary), and can be referenced by any function of the agent (including Inner Agents and their functions)
• This can be useful for non-text modalities (e.g. audio, pdfs, image) and lengthy text modalities, which we do not want to output into subtasks_completed directly
• To use, simply define an External Function with shared_variables as the first input variable, from which you can access and modify shared_variables directly
• The agent will also be able to be self-referenced in the External Function via shared_variables['agent'], so you can change the agent's internal parameters via shared_variables
• If the function has no output because the output is stored in shared_variables, the default return value will be {'Status': 'Completed'}

# Use shared_variables as input to your external function to access and modify the shared variables
def generate_quotes(shared_variables, number_of_quotes: int, category: str):
    ''' Generates number_of_quotes quotes about category '''
    # Retrieve from shared variables
    my_quote_list = shared_variables['Quote List']
    
    # Generate the quotes
    res = strict_json(system_prompt = f'''Generate {number_of_quotes} sentences about {category}. 
Do them in the format "<Quote> - <Person>", e.g. "The way to get started is to quit talking and begin doing. - Walt Disney"
Ensure your quotes contain only ' within the quote, and are enclosed by " ''',
                      user_prompt = '',
                      output_format = {'Quote List': f'list of {number_of_quotes} quotes, type: List[str]'},
                      llm = llm)
    
    my_quote_list.extend([f'Category: {category}. '+ x for x in res['Quote List']])
    
    # Store back to shared variables
    shared_variables['Quote List'] = my_quote_list

• Global Context is a very powerful feature in TaskGen, as it allows the Agent to be updated with the latest environmental state before every decision it makes

• It also allows for learnings in shared_variables to be carried across tasks, making the Agent teachable and learn through experiences

• A recommended practice is to always store the learnings of the Agent during the External Function call, and reset the Agent after each task, so that subtasks_completed will be as short as possible to avoid confusion to the Agent

• There are two ways to use Global Context, and both can be used concurrently:

  • 1. global_context
    • If all you need in the global context is shared_variables without any modification to it, then you can use global_context
    • global_context is a string with <shared_variables_name> enclosed with <>. These <> will be replaced with the actual variable in shared_variables
  • 2. get_global_context
    • get_global_context is a function that takes in the agent's internal parameters (self) and outputs a string to the LLM to append to the prompts of any LLM-based calls internally, e.g. get_next_subtask, use_llm, reply_to_user
  • You have full flexibility to access anything the agent knows and process the shared_variables as required and configure a global prompt to the agent

• We can use Global Context to keep track of inventory state
• We simply get the functions add_item_to_inventory and remove_item_from_inventory to modify the shared_variable named Inventory
• Note we can also put rule-based checks like checking if item is in inventory before removing inside the function
• Even after task reset, the Agent still knows the inventory because of Global Context

def add_item_to_inventory(shared_variables, item: str) -> str:
    ''' Adds item to inventory, and returns outcome of action '''
    shared_variables['Inventory'].append(item)
    return f'{item} successfully added to Inventory'
    
def remove_item_from_inventory(shared_variables, item: str) -> str:
    ''' Removes item from inventory and returns outcome of action '''
    if item in shared_variables['Inventory']:
        shared_variables['Inventory'].remove(item)
        return f'{item} successfully removed from Inventory'
    else:
        return f'{item} not found in Inventory, unable to remove'
    
agent = Agent('Inventory Manager', 
              'Adds and removes items in Inventory. Only able to remove items if present in Inventory',
              shared_variables = {'Inventory': []},
              global_context = 'Inventory: <Inventory>', # Add in Global Context here with shared_variables Inventory
              llm = llm).assign_functions([add_item_to_inventory, remove_item_from_inventory])

• There are other features like Memory (Tutorial 3), Hierarchical Agents (Tutorial 4), CodeGen and External Function Interfacing (Tutorial 5), Conversation Class (Tutorial 6)
• These extend the baseline features of TaskGen and you are encouraged to take a look at the Tutorials for more information.

• gpt-3.5-turbo is not that great with mathematical functions for Agents. Use gpt-4o-mini or better for more consistent results
• gpt-3.5-turbo is not that great with Memory (Tutorial 3). Use gpt-4o-mini or better for more consistent results

3. Other Known Limitations - Do test the framework out extensively and note its failure cases. We will see if we can address them, if not we will put them in Known Limitations.
4. (For the prompt engineer). If you could find a better way to make the prompts work, let us know directly - we do need to test this out across all Tutorial Jupyter Notebooks to make sure that it really works with existing datasets. Also, if you are using other LLMs beside OpenAI, and find the prompts do not work as well - try to rejig your own prompts and let us know as well!