Managing state across multi-step agent workflows in LlamaIndex — patterns?

[Arpanwanwe]

I have been using LlamaIndex for agent-like workflows-planning, execution, tool usage, and sometimes reviews. One thing I keep running into is state management once the workflow sprawls to multiple steps.

In other words, when agents hand work off to each other, context ends up split between prompts, intermediate nodes, and application code. This gets especially tricky with retries or branching logic-it's hard to tell why an agent made a certain decision or to reproduce behavior later.

I have tried to implement this by keeping an explicit shared specification/state that agents read and write, instead of relying solely on the implicit prompt context. I've been testing this with a small orchestration tool, Zenflow, in conjunction with LlamaIndex to make workflows more inspectable, but I'm still feeling out whether this is the right direction.

Wondering how people here are handling this with LlamaIndex: mostly manually managing state, using callback handlers/state handlers, or leaning on the newer abstractions for agents in the library?