Agentic AI at first glance sounds like an unsurmountable project. How do we create this always-helpful, proactive assistant that knows it all and knows all about me? The answer, as so often, lies in breaking things down into more manageable pieces and then putting them back together to form that powerful system. Looking at it this way makes one quickly realize that a nice thing about Agentic AI is that many of the intermediate steps already exist in various forms and can be used as initial building blocks.
But what are those building blocks in an agentic setup? At its core, Agentic AI consists of three main pieces: tools, workflows, and then those ominous agents. But here we’re going to make it four, because in the context of GenAI, it is useful to distinguish between two different types of tools: tools that use more traditional techniques and tools that make use of GenAI.
Tools, Workflows, and Agents
Tools get stuff done. This can be as simple as aggregating some data, accessing an external data source, classifying an object, or predicting future events. Those are often tasks that we have solved via various—often automated—tools already in the past.
However, note that for Agentic AI, it is much more useful to be provided with content and not just access to a data source. So, for example, instead of creating a tool that retrieves data from a CRM, it is more useful to create a tool that retrieves information about a customer (or customers). That tool might still simply access the CRM under the hood, but it’s more likely to aggregate data from a few data sources. Similarly it’s more useful if a tool could be providing more information about users preferences rather than just giving access to their raw calendar data.
Intelligent Tools get stuff done by making use of GenAI.
Some call this augmenting analytics with AI. A typical example could be tools that translate or summarize a document or a tool that predicts the sentiment of a message. Those are typically the types of tools that are now a lot easier to implement than before, when e.g. document analysis was a science of its own. Another example could be a tool that checks images, for example to be sure that an infographic is consistent with corporate style guidelines.
Nowadays there is a risk of turning every tool into an intelligent tool even though classic techniques can solve particular problems as easily (and often faster and cheaper). Avoid falling into the “everything looks like a nail if I have a hammer”-trap.
AI Workflows line up a series of tools to solve more complex tasks.
This could be a series of calls to:
- First retrieve information about a customer using a data aggregation tool, then
- Summarize this information using an intelligent text summarization tool,
- Translate it to the account manager's native language and finally
- Send this information via email to said account manager.
It could also be a tool that automatically:
- Analyzes new publicly published material,
- Checks it for style and tonality guideline violations and
- Informs the author about those issues with suggestions for improvement.
Note that an AI workflow can, itself, become an (intelligent) tool used by other AI workflows.
[Side note: From a KNIME perspective this is potentially a tad confusing. AI workflows focus on orchestrating the way tools are lined up. A KNIME workflow does a lot more, it can not only control tool usage but also model data and other control flows around it.]
Agents: this is the step where it gets interesting - or truly agentic. An agent has access to a (potentially large) repository of tools and orchestrates their use differently for each query.
How this is done under the hood depends on the actual implementation. The agent could first use an AI to create a plan, which includes the set and order of tools to be used. But the agent could also make use of tools iteratively, until it has enough information/material to respond or take action.
Agents come in two flavors:
- Agentic Applications that can interact directly with the user or
- Agentic Services that run in the background and can also be used by other AI workflows or agents as a tool.
The true power of this seemingly simple setup comes from the ability to add new tools to the repository that are under the hood, AI workflows or agents. These new tools can be added by human designers but they can also stem from agentic work itself, that is agents add more tools to their own or other agents’ toolboxes.
The more complex the tools in that repository become, the more complex tasks can be done by the agents using them. So over time, more complex behaviour arises naturally. Remember the old divide and conquer? The small pieces aren’t all that complicated but what comes out when they play together can be!
Tools, AI Workflows, and Agents in KNIME
From a KNIME perspective, all of those components can be built naturally using KNIME workflows:
- Tools: Creating tools, especially ones that make use of data, is one of the core capabilities of KNIME workflows.
- AI workflows: Lining up tool calls, or building AI workflows, is a strength of the platform. That way a space on the KNIME Hub can turn into a repository of a variety of different tools - some also intelligent using the AI extensions that constantly add capabilities when new AI models or methods are created.
- Agents: Using those AI capabilities also enables KNIME users to build agents by adding reasoning prompts to generate plans or prompts that ask the AI to refer actions to some of those provided tools.
KNIME workflows can also be used to build agentic applications and services. Just as normal workflows can be deployed as applications using data apps or services using endpoint deployments, an agentic workflow can be deployed on the Hub as well.
If you're interested in Agentic AI, I'll be discussing it further in my talk at KNIME Spring Summit in Berlin, March 18-20, 2025.