At the Moment development is in a jupiter file named ProtoOne.
The Agent is based on a more sophisticated Task Pool concept. The Tasks are connected in a Graph. There a Parent and Child Tasks... more info in the notebook
Agent recives guidens by human user in planning and execution phase
Agent only recives guidence in the planning phase
Agent searches autonome to improvments of his problemsolving capapilities and tests them
Allways perform a langchsin compress retriever step over available information.
Task tree should be at the heart.
Smallest unit of work: Compress->select tools->think->improve->act commands->describe current state
Task description and prompts in the eli5 format. But use 12 or so
Add curent list of fils allways in problem solver promp with their stats like pagecount
Add local CoT and reflection step in solver
Tell it that looking at file means iterating over it
Split function dissolves main task. Maybe call function that creates result description when all child's done
Add curent list of fils allways in problem solver promp with their stats like pagecount
create a Task complexity evaluation (maybe arround number of expected commands needed) and use this to split tasks
agent needs a portfolio of suggested workflows that help to handle the small contextsize. explain how to read big files page by page write novels chapter by chapter and keep a file containing current state of the novel(where wich person is and so on) that changes after every new text written
remove oldest messages when contextwindow gets full in task solver
Tell the Tasksolver how much Context s used, and if hes in danger, to loose context
add the scratchpad to the solver function
Let the File Operations return better error messages and tipps.
The agent gets distracted when he can send comands. make better signal, that its time to send command
It "copies" a file but halucinates extra contend into it