hey guys I am stuck at the logic or appraoch the way to create my autoamtion.
Describe the question
Imagine, I have a prompt : create a meeting with David on wendesday 11 AM.
expected action is : checking calendar both on me and David and looking if empty spot is matched. if not it issupsed to suggest (find) another match and suggest. after that it is supposed to drop a message to David on telegram asking him if you are okay with this meeting. If he accepts that, then meeting will be set on calnedar and also trello.
What is the appraoch to tackle this request?
One appraoch is to let AI Agent design the sequence and use the tools in order to handle the reuest. E.g., lookup in database (user_tool) who is David (fetching relavant metrics such as Trello ID and so on) . lookup in calendars of both (calendar_tool), get confirmation on meeting from david (telegram_tool) , setting the meeting on calendar and trello (calendar_tool & trello_tool)
or I make sub-workflow as tool in which receive this high level request and handling the request through some nodes in its flow. e..g., use a firebase node to get info on david , then use a googel calendar node to get empty spots for both , then use telegram node to drop message on david , and so on
Pros & Cons
I think the second apprach use less of LLMs, thus reducing the cost spending on AI. let the flow handle all complicated lgoic. avoid mistakes by LLMs.
the First appraoch on the other hand seems more professional and easier to expand. however it is costly and I also have issue with loading many prompts into AI Agent as system is gorwing.
Solution
what are you thoughs on this ? any experience on big scale project ? can provide me example of smilard to my case.
I don’t have examples unfortunately, but it sounds like cost is a big decision driver here. That alone might be enough to switch to nodes over AI, but I personally also prefer more structure over more AI, so your second approach sounds better in my opinion.
I would value the tool working reliably over how professional it looks, unless the AI version can be tested extensively so confidence in it is high.
it looks old fashion way .e.g, using classification and then handling all matter ina a workflow ? then AI Agent is somehows usless. Think of a human who has a brain and some functionality thorugh hands and legs. Brain knows what to do with hands and leg to get the job done. It is not like Brain says I know hand what capable of , i will pass the request to hand and it will be okay . hands know what to do.
I really don’t think you should care how it looks or feels - the workflow is there to perform a task and the only thing that matters is how successful it is.
Overall, both methods are ok, I would definitely prefer the AI one but only if it’s reliable enough. This can only be tested by you, according to your needs and expectations, of course.
this is straght forward. there is only one agent and one platform to be engaged. imagine you have 10 platforms and each have 10 tools to be exectued. so it is not going to be this easy. also think of the size for prompt it would be huge if you are supposed to use only AI Agent with 100 tools in one place.
I see, didn’t really catch your point though. The way I see it, if you go down the AI path, you could split tools across multiple agents - 1 AI agent with 100 tools and a huge prompt is a recipe for disaster IMO.
