Making sense of AI agents and Agentic AI
Not a day goes by without new agent announcements.
Cutting through the hype to discern what’s truly offered and genuinely new is becoming increasingly difficult, and I’ve been riffing on a high-level framework for some time.
The recent taxonomy paper by Ranjan Sapkota, Konstantinos Roumeliotis, and Manoj Karkee helped me settle and sharpen several key ideas.
Their research highlights the differences between AI Agents and Agentic AI, a critical nuance often overlooked. Two key factors set Agentic AI apart.
First is scope: AI agents focus on specific tasks, while Agentic AI tackles broader jobs-to-be-done (JTBDs) by breaking them down into multi-step workflows.
Second is orchestration: Agentic AI doesn’t just act—it coordinates. It enables multiple agents to collaborate, share context, and adapt to drive more complex outcomes.
I believe it’s important to stress that the AI game isn’t a beauty contest—there’s real value in being purpose-built for a specific task.
Digging deeper into agents, I find it useful to categorize them into two types:
Ingress agents are embedded in existing software to simplify task execution. They eliminate the need to navigate complex screens. They can provide a human-like conversational interface or be triggered by events. They seldom change what the application can do but transform how it’s done, pushing the envelope of usage and adoption. They can become transformative when they gain access to broader data sets.
Egress agents operate within applications, invoked by workflows via MCPs or other mechanisms. They extend application capabilities, streamline cross-application workflows, and push the boundaries of automation without human intervention.
Reasoning is the hardest capability to assess. It depends on the scope of the decision, which can range from a simple task to guiding and resolving an entire job to be done. At this stage, I’m also finding it can be a distraction: AI Agents and Agentic AI already offer significant value, and most initial CX and Sales use cases will be driven by prescriptive workflows.
I would love to get your feedback as I begin updating my CX and SalesTech industry landscapes: diverse perspectives have consistently proven invaluable!