Hey everyone. As many, I have been digging into the recent developer releases: AgentKit + ChatKit.
AgentBuilder looks great: a great medium between a developer and other departments.
However, I’m confused with its usability between me (dev) and whoever else is organizing the visual side of things. Since a lot of the code needs to be re-written (Approvals, Tools, etc) how am I supposed to streamline this?
For the Agents SDK it looks… okay; I can come in and fill in the blanks. Yet with the ChatKit version it (AFAIK) is stuck to using the workflow ID.
So, realistically, if a department worker publishes (because we are not opting for preview) the workflow and doesn’t understand that I still need to fill in the blanks, it will break the output. This feels fundamentally flawed and leads me to worry that - like many OpenAI releases - it will be left to collect dust.
Also, side note: Why are we required to verify to preview? I wouldn’t in a million years trust any sort of third-party verification tool - especially when the reasoning behind it seems arbitrary.
The only reason I can see myself - and my team(s) - using this is because it’s free. Otherwise it feels disjointed and improperly implemented.
Last note: The topics here need to be updated; there is nothing to select for AgentBuilder. (Thank you for updating the tags)
My thoughts are that it is more than novelty, but it feels like a work in progress.
For now, I can see it only being used as a kind of prototyping tool for agentic development but you need to take the code AgentSDK code that is generated and add to it to get something useful.
Exactly, which disjoints any sort of collaborative experience. I have been messing around with it & ChatKit (which is even less control than using the AgentSDK). Although very fast to deploy, I’m having a difficult time trying to understand how to use it productively with a team.
I think the collaborative aspect of the experience is working in a visual paradigm with business users and developer(s) together and being able to prototype and evaluate the automation of a business process. Once it is validated, taking the AgentSDK code and expanding on it. That said, I am finding that there some things that just don’t work as they should just yet. File Search or MCP do not seem to work as a node only as a tool of an agent. I haven’t found a way to have agents running in parallel yet etc. The conditional node seems to work reasonably well but context gets lost coming out of it requiring you to define and set state variables. I think these shortcomings will get ironed out in time, but it is the reason I said this feels like a work in progress.
It’s a great concept and might be super useful for companies in the future. For now the amount of options are very limited and not great at all for projects that are still in development stage, from my experience in the last few days.