Problem Statement

Currently when using the agent_graph tool, you don't have full control over what parameters are passed to the use_llm tool call to evaluate your graph. This is problematic as the default callback_handler prints the agent output to stdout. If you have a graph with multiple concurrent ages this leads to colliding output on stdout and the results are unreadable. I would like to be able to override the parameters sent to the use_llm call such that I am in control over each agent's configuration.

Proposed Solution

The agent_graph function does take kwargs, from which some values are added to tool_context which is eventually passed to the use_llm function, however since we don't take all values from kwargs, just the known ones, then I can't override arbitrary parameters with this. A naive solution would be to pass all of kwargs to tool_context.

However, this doesn't work nicely if you want different agents to have different parameters. This feels like this would have to be assigned as part of the graph topology, but it's a larger change.

Use Case

agent_graph is almost unusable from a stdout perspective without this change.

Alternatives Solutions

I may be able to hack this by interacting directly with the _MANAGER variable in the agent_graph file, but this feels flimsy.

Additional Context

No response