I wanted to clarify the behavior of Retool Agents in multi-turn interactions.
From what I understand:
Tool outputs are persisted between turns, but
They are not automatically included in the input context of subsequent turns, only previous agent responses and user messages are passed forward.
This means the agent needs to re-run the same tools in each turn to regain access to their outputs, which adds latency, complexity, and of course cost.
Is the above an accurate representation of current state?
And if so, is there any plan or timeline to enable tool results to be passed automatically into subsequent turns?
Hey @MiguelOrtiz thanks for the bump & sorry for missing this post.
You're correct about the current behavior, which was intended to be a context-preserving decision, but I agree that in multi-turn chats it's suboptimal.
Right now, our focus is on self-hosting agents, source-controlling agents, adding OAuth to MCP tools, building synchronous agent calls from workflows (to remove the need for polling), and making it easier to add eval dataset items, but as we wrap this work up in the coming weeks, this is among the most important things we'll tackle next. From a timeline perspective, I'd expect it around late September or early October.