Currently the agentic flow is something like:
Task + (Optional) Context -> (Optional) Reasoning + Code
Task employs both the Goal, but also the Observation (together with context) in the sense of a traditional RL agent. I think if we change the flow to
Observation(s) + Goals -> Reasoning + Goals` + Action(s)
Where Observation(s) can consist of the Task and Context, Goals and Goals` are the goals of the agent before and after the turn, and Action(s) are the code block, it could be more general purpose than one flow step by step task solving agents, which the repo currently provides. I'm opening this issue to facilitate discussion, and it will stay open only until after dedicated issues have been opened and solved on this matter, everyone's opinion is welcome.
As a one step improvement, the first update could look like:
Observation -> Reasoning + Actions
and then it could be built upon to integrate Goals.
Currently the agentic flow is something like:
Task employs both the Goal, but also the Observation (together with context) in the sense of a traditional RL agent. I think if we change the flow to
Where Observation(s) can consist of the Task and Context, Goals and Goals` are the goals of the agent before and after the turn, and Action(s) are the code block, it could be more general purpose than one flow step by step task solving agents, which the repo currently provides. I'm opening this issue to facilitate discussion, and it will stay open only until after dedicated issues have been opened and solved on this matter, everyone's opinion is welcome.
As a one step improvement, the first update could look like:
and then it could be built upon to integrate Goals.