I’m building an AgenticSearch with recursive output improvement system using the OpenAI Agent SDK and am deciding between a CLI-focused interface and a webpage interface. I’ve also added secure session with login and 2FA authentication, and sub-agents that synthesize the search results into a concise final output. For anyone who has deployed similar agent workflows, which approach ultimately offered better usability and faster iteration? Did the added frontend complexity pay off, or did the CLI remain the more practical option in real-world use?
1 Like
Perhaps I could use both a CLI and a webpage interface, but I’m not entirely sure yet. What are the trade-offs between a CLI and a webpage?
Hi, happens with me as well. Any idea?
Welcome to the community!
I consider my interface one of my personal best achievements… It started in CLI and moved to browser and GIF… Simple fractal designs will naturally work best with AI I believe as it scales well… In fact I believe the dividing tree I use to be the most efficient ‘AI interface’…
I believe that effective AI coding means combining many abstract concepts so labelling those concepts and using those names as the structure of the tree you remove any wastage of space on the page, making the page a tree view. I hinge trees with iframes with each tree as a module grouped into forests…
I even build all my tools in it with an Edit mode so a fully editable modular structure which acts as an interface. Every ‘pane’ or branch is a new decision available to take.
I have a constantly running task manager ‘Phoenix’ to run/manage module set tasks - I’m still considering whether best through CLI or Web UI structurally too for this ‘Orchestrator’.
I know it is confusing everyone here speaks a different language to me (and I speak English) so consider my structure as:
Forest → Like a folder for a concept
Trees → A constructed code tree (not sure it’s quite AST exactly but that general idea)
Tree Branches → Visualized code processes
PhasmIDs → These are the agents, themselves Tree Modules
This way you have a nice ‘fractal’ modular design.
I demo it on the forum in all my projects.
I think it’s not smart to be designing UI when we have AI… There is likely little scope that hasn’t been trialled in web browser, it is a well established and multi-modal high resolution area. We don’t need interfaces for everything… Why do you need an interface… I think OpenAI are suggesting completely interface free device?.. Well I believe you do need an interface sometimes and so i am chasing the best generic AI design 
.
If by ‘real world use’ you mean production then I don’t do that… I see huge AI space like fresh snow and want to be everywhere at once.
That said… I do imagine this is at least one of the most efficient multimodal Web Browser designs for AI? I would love to understand better
My advice for web browser interface for what it’s worth… Get a really good replicating base structure and get AI to make the nodes appropriately beautiful…
I always used to build forms into webpages, it’s just the same adding JSON endpoint return too… With a editable self replicating modular interface you have everything from a few lines of code + the core AI Orchestrator or ‘Kernel’ in my OS terms.
I started in CLI with menus, then executing code either side of menu execution etc etc. One of my first ‘AI’ features to have the menu reshape based on last usage time before any of todays AI algos
I think CLI maybe better with AI device screen restrictions, certainly this is my focus on Tree structures and menus… Compressing as much data into the screen as possible through both data and structural choices.
I tend to associate CLI with Batch and Web UI as real user interaction.
My design might not be good for ‘production’ I don’t really know… My goal is just the best generic AI UI design.
I have done some work on web indexing, if you are doing recursion THAT is what you are watching with this interface… You want managed recursion within your modules I think rather than recursive modules to prevent overflow.
Oh and my Trees have authentication in some versions so any tree can be secured with for example user/group…
in your case each sub agent you might represent as a Forest, Trees you would use as different views on those Agents, PhasmIDs would be the (complex not primitive) processes they run.
Hope it helps
1 Like
Thanks for the help. I really appreciate it.
1 Like