ParthSareen / DAGent

AI Agents as DAGs

DAGent - Directed Acyclic Graphs (DAGs) as AI Agents

How to Use

Install the lib

pip install dagent

DAGent basics

DecisionNode

• This is where the llm picks a function to run from given options
• The .compile() method autogenerates and saves tool descriptions run with param force_load=True if there are errors or if an option of tool changes

FunctionNode

• Runs a python function
• Can be attached to a DecisionNode to be treated as a tool and allow an LLM to choose what to run

Dagent Diagram

Example Usage

See dagent/examples/simple_agent.py for more info

1. Put all functions to be run into nodes

def add_two_nums(a: int, b: int) -> int:
    return a + b

add_two_nums_node = FunctionNode(func=add_two_nums)

2. If you want any decision making steps, each needs to be a node

decision_node1 = DecisionNode()
# decision_node2 = DecisionNode()

2. Link the appropriate nodes together

The compile method will link it appropriately

decision_node1.next_nodes = [
    add_two_nums_node,
    multiply_two_nums_node,
]

3. Run the .compile() method

• Make sure to run this on the first function, the rest will get compiled as well
• It will save json for the tools it has compiled for the attached functions

decision_node1.compile()

4. Run the entry function

decision_node1.run()

Other things to know

• prev_output is needed in the function signature if you want to use the value from the prior function's value. Obviously the prior function should have returned something for this to work
• Args can be overriden at any time using the following (this merges the kwargs in the background with priority to the user):

add_two_nums_node.user_params = {
    a : 10
}

Motivation

• I found it difficult to use existing libraries and spend the extra time learning a framework to build an "agent"
• Was spending too much time reading their docs and writing agents manually was just faster
• So I built a framework to deal with all the things I didn't like about agents and help me structure code to be build agents quickly

Things I am looking into (I have lots of opinions)

• Use "layers" to track next nodes instead of linkedlist style - not gonna do this prematurely :)
• Probably need a way to return a value directly from a top level run function
• Look if things run in memory and how to isolate for large workflows -> e.g. funcA(funcB(...)) -> funcA(...) -> funcB(...)
• Side effects/mutations
• Creating a data model for communication between functions + schema validation -> autogenerate?
• Logging
• Alerting on error
• Add a compile method to derive data models and tool descriptions

Acks

Shoutout to:

• @omkizzy
• @kaelan