eigent-ai
[Feature Request] Task Agent Planning Improvements: Environment Awareness, Tool Selection, and Replanning Strategy
Motivation
- When the agent lacks environmental understanding during initial task planning, the plan is likely to be wrong.
- We need to clarify: when should the agent execute first and then re-plan, and when should it gather more information before planning?
Solution
@Wendong-Fan
-
Provide the task agent with a terminal and a browser. When facing a directed but unfamiliar environment, it should first browse or gather environmental information before making a task plan.
-
Set up four tools for the task agent:
- Directly respond (when the task is relatively simple).
- Go directly into workforce planning (when the task is complex and the environment is unfamiliar but not directed).
- Obtain environmental information first, then enter workforce planning (when the task is complex and the environment is unfamiliar but directed).
- The task agent chooses which tool to invoke based on the current task.
- When a task goes directly into workforce planning, a fault-tolerance and retry mechanism needs to be designed. That is, when the agent makes mistakes but obtains more useful information, it should be able to re-plan the task.
Alternatives
No response
Additional context
No response
@Douglasymlai
From UX side (pending) :
We need to design the chatbox flow for the following three cases:
Directly respond → when the task is relatively simple.
Go directly into workforce planning → when the task is complex and the environment is unfamiliar but not directed.
Obtain environmental information first, then enter workforce planning → when the task is complex and the environment is unfamiliar but directed.
The key update is that the task agent now has the ability to call both the browser and the terminal.
Open Questions
Should we visualize the task agent’s workspace?
If not visualized, the MVP version could display the task agent’s steps in the subtasks box after the algorithm finishes.
For example, print logs such as:
"Calling browser_toolkits… fetching IT ticket submission website content..."
"Gathering environment information, starting to plan the task…"
