Suggestions on how to go about dynamic code generation using multi-agent approach

[invalidexplorer]

Hi everyone! I need help with finding an approach to a particular problem. I am using React agents to utilize Python to open different CSVs (large sizes that exceed the context window) and understand their schemas. Then, according to some external rules fed into the agent as text, I make the React agent with PythonREPL as a tool to write code that queries these CSVs, maps them to each other, and creates rule-based filters and output CSVs after all of these different operations.

What I noticed was that the output given by one React agent was inconsistent, although sometimes it was correct. To improve its consistency, I used this example: 'https://github.com/langchain-ai/langgraph/blob/main/examples/plan-and-execute/plan-and-execute.ipynb' with the React agent being the executor.

However, the problem I'm facing is that after the planner comes up with steps and makes the React agent execute each step, the context generated like the thought process in the scratchpad and other things are shared in a limited manner with the agent executing the task. If the agent depends on things like a DataFrame generated during the previous iteration, the executor agent is unable to complete certain tasks. It even begins assuming sample data (Despite it being promoted not to) to complete the tasks.

How do I approach dynamic code generation for a project that involves multiple complex steps in Python? (The Alphacodium example might not suit my needs as there are multiple steps involve, some involve opening the csv and observing certain values for the next iteration).

[sankar-masori]

@invalidexplorer : Did you solve the issue? currently I am facing similar problem. Can you please share some pointers?