# Agent toolboxes should they be inteligent? **URL:** \ **Category:** Research\ **Created:** [October 2, 2024, 6:40am UTC](https://discuss.huggingface.co/t/agent-toolboxes-should-they-be-inteligent/109758 "2024-10-02T06:40:27Z")\ **Posts on this page:** 1\ **Page:** 1
**Author:** ![LeroyDyer](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/leroydyer/32/26511_2.png) [@LeroyDyer](https://discuss.huggingface.co/u/LeroyDyer)\ **Post date:** [October 2, 2024, 6:40am UTC](https://discuss.huggingface.co/t/agent-toolboxes-should-they-be-inteligent/109758/1 "2024-10-02T06:40:27Z")
Here i have found that if you give a oneshot or multi shot example with the current tools it forces the model to use the data in the prompt ? even if it is false ! … So A fake tool instead of a real tool ! as the example is only to show the model the WORKFLOW ! not how to call a tool ! In model which have not been trained for tool use in that instance : you will need to specify how the tool use is expected ( i personally use docstrings ) … ( adding the docstrigs for the agents tools to the prompt somewhere ) and in others i use the standard openAI prompt messaging pyantic template or i even use Inspector ! and other iu use Hand crafted Extraction methods ! as eAch methodolgy in the function calling or structured response and workflow has its advantage and disadvantages ! In some models i extract all functions from the response to files ! as well as execute the fragments : and provide the outputs to the fial response : as the final response can also be formeted by the api removing that from the model ! enabling for faster execution ! hence a good process should offload some work to the tool ! and use the tool to return the data rather that using the model to do all the processing of information ( in some cases ) … Personaly i think of the model as a Personality on top of a library of DATA ! so Assitance fro this agent is in the form of tools ! We discover that we cannot actually give the model a whole list of tools ! its a problem ! but we can keep a rough list of tools ! or a graph of tools ! or a folder of tools ! … in which the model can search the list for potential tools , then load the docstrings for the tools ! and use them and dispose of the tools after use ! even to generate a tool or chain instead 1 saving the tool to the folder and adding it to the available tools list dictionary ! most imortant tools are NAMED as thier useage ! the model can even Guess its usage ! hence picking potential rools ! then getting thier args ! SO it would kow to create missing tools ! … Ie roles demand tools from tool creation agents ! So if a role needs a tool the role can request it ! currently i am working on the interactive tool box ! with its own agent ! ( like rogue squadren ) the solder carrys a ai in his gun and one in his back pack ! so if he needs some thing the backpack will know before him ! LOL ! but this is the way forwards ! a toolbox with its own agent serving tools to the agents or creating themm as required and managing the tool box !