I’d love to see multi-agent made easy within the low-code environment KNIME provides.
I have experimented a lot with the Agent functionality and it works AMAZING - good job to everyone involved to make it as simple as it is today - I am not sure if there is any other tool / framework that is as easily accessible and as customizable as what KNIME offers in this space.
I do believe that this can be made even better, by adding support for multi-agent - let me share some of my experiences (some of you may have been in Stuttgart when I did the live demo…):
- Agents work really well for process tasks - even cross system
- There seems to be a reliability limit once ~20 Tools are exceeded for a single Agent. I am not sure if this is solely due to complexity of instructions in prompts, Tool description, config node descriptions etc and simplifying that would mean this ceiling “rises”, but let’s just for no go with there being some sort of hard limit - I will assume in the text below that this is 20.
- With this hard limit, scaling Agents using KNIME right now means:
- Creating specialists in certain tasks that have a maximum of ~20 Tools
- These specialists will need to be accessed using different data apps (or in the same data app with a “specialist” selection that activates the respective branch
- I do believe this is inconvenient and leads to sub-optimal experience of the end user / agent consumer as they will have to keep track of when to use which agent
- There is also a risk that this will limit moving end-2-end processes to a single agent, if more than 20 tools are required to perform this
- In order to circumvent this hard limit I have experimented with the following multi-agent set up:
- One entry point for a user to a Front Desk Agent / Orchestrator agent
- The tools of this agent are sub-agent (specialists) (Agent Prompter + Tools) which are available to the Front Desk agent
- Due to limitations of how relative paths resolve etc. the set up to give the sub-agents access to tools is messy and complicated to maintain (although you may guess it - I have a workflow for making that a bit easier…)
- The sub-agents manage a specific part of a larger process and have the required tools
- The sub-agents have instructions / knowledge about the process sequence and what information is required by which other sub-agent.
- Right now in order to maintain this in a simple manner I actually am running a single agent set up with all tools in parallel to a multi agent set up and have a process in place to update tools for multi agent setting
- The results of Multi agent are really good in my view and this seems to, in my view, be a great way to implement the vision of potentially having one point of entry to a companies agent (Front Desk) to then access a whole bunch of agents that can take care of complex tasks - let’s just assume that 20 is the hard limit - imaging Front Desk having access to 20 sub agents and each have 20 tools - that’s 400 tools being accessible - and that is at nested tree depth = 1…
The multi agent set up by simply nesting sub-agents as tools of front desk agent seems to be somewhat intuitive and should be very easy to grasp by the average citizen developer that wants to get into development of agentic systems - I think it resonates really well, because it somehow feels like an org chart… so my first wish would be: Solve the path problem and allow people to build nested agents (and maintain them) in an easy way
I’m sure that is not without complication to resolve, but in my simple mind feels like the easiest alternative
Other options could be to look into A2A (which was actually mentioned by one of you in Stuttgart) although I am not yet to familiar how that set up would look like (but will do some research…)
Anyways - I feel that with what certainly is to come in the space of Agentic AI - you guys have the opportunity to strongly position yourself - KNIME connects to pretty much everything so let all the other companies build their vendor-dependent systems inside their solutions and you focus on being the spider in the middle to connect them all ![]()
Happy to discuss my set up and experience in more detail….