Agentive Tech and Agentic AI: Cousins not adversaries
In 2017 Rosenfeld Media published my book Designing Agentive Technology: AI That Works for People, all about the issues involved in designing tech that does work on users’ behalf while their attention is elsewhere. It’s seven years old now, but I’m glad to report it seems to have stood up, even if the infrastructure to support such designs was not commonly available at the time of writing.
Then, over the past several years, we’ve seen AI luminaries including Andrew Ng speak about “Agentic AI” and lots of folks have reached out to me to help make sense of whether and how these two things relate.
It’s an understandable request because “agent” is baked right into both names. But, like “baby powder”, “baby food”, and “baby back ribs” their shared referent does not necessarily mean they are the same thing. The rest of this article explains why and how.
Side-note language pedantry
As I’ve noted elsewhere on social media, my language-nerd hackles are raised by the word “agentic”, since “agent” is a Latin-based word, and the Latin suffix for adjectived-noun is -ive, whereas -ic is Greek. That’s why we would write “festive” rather than “festic” for the Latin-based word “festival”, or “geographic” rather than “geographive” for the Greek-based word “geography.”
But we shouldn’t get too bent out of shape about it, since there are other words like this that go unchallenged in English, like “artistic” (Latin root artis) or “passive” (Greek root πάθος). Language weird.
Moreover, there’s still a good reason for keeping “agentic” around—it’s because they’re different things.
Teasing them apart
Yes, both of these concepts involve agents, acting on behalf of a user. The main difference is that the agentic framework is a “back stage” thing — being about how a user request is handled in the code — and the agentive framework is a “front stage” thing — being largely about the user experience implications of a thing being handled out of the user’s attention.
Agentic?
Going into a bit more detail, agentic AI is a way to take a natural language user request, use large language models to break it into an actionable plan, and have each step handled by a software agent that is good at that particular thing; like web search or handling math, writing code or looking for potential harms, improving output through Socratic challenges or formatting the final output. Agentic frameworks enable a user to ask for things like “List the nouns that have appeared uniquely in the dialogue of every movie that has won Best Picture since the advent of the Oscars” and have a system start to work on it. Current large language models kind of panic about such requests.
Ng also talks about how agentic systems may provide a first draft answer and then refine it over time as steps are completed and the answer expanded or refined. One imagines subscribing to the output like a wikipedia article to be alerted when those updates occur. (In a wholly different context I described this as an “active academy” model that contrasts a “stoic guru” one, if you find that clarifying.)
Agentive?
In contrast, the agentive framework is a “frontstage” thing. In the book I work to make this concept distinguishable, and list the unique use cases and questions that a designer may need to account for as they work on such a system. How does a user set up and refine their preferences? What’s the handoff and takeback use cases? How should users decommission an agent when it’s no longer useful? In good pedagogical fashion, I focused on the basic case of products that are strongly agentive.
In talks, articles, podcasts, and workshops, I’ve expanded on the bigger picture of how agents stand to outperform the experience economy, how the modes of agency fit together, how mature products and features often span these modes, risks and mitigations, and how agents differ from assistants and how the shifts between them can happen effectively.
One way to think about the difference is to ask, for each framework, who is boss of the agents? In the case of agentic AI, the agents are working on behalf of the primary AI, the thing coordinating all of their efforts and consolidating them for the user. In agentive technology, the agents are working on behalf of the person, the user. It’s why I’ve occasionally framed the adoption of agentive tech as giving users a promotion.
They’re complementary, but not the same
These two frameworks are quite related. Is it easy to imagine a user making a request of an agentic system and having that experience informed by agentive designs. But they are not inextricably intertwined.
Being a technology architecture, an agentic AI could bypass the need for agentive concerns. “What is 1+1?” should get an instant response and only need assistant guardrails. In the future, when computers are orders of magnitude faster than they are now, you could imagine an agentic system handling much more complicated questions, but still doing so instantaneously.
Being a user experience framework, agentive tech can be driven by non-agentic AI. A quick example is an alarm clock. In fact, since I published the book before the rise of foundation models, all 47 examples I cite there — like the Roomba vacuum cleaner or the CatGenie litter box — are all agentive despite not being powered by agentic AI.
So sure, though “agentic” sounds like nails on a chalkboard for language reasons, it’s useful to have different words to distinguish the front-end strategy and UX questions of agentive technology from the largely back-end technology structures offered in agentic AI. One needn’t “win out” over the other. They are cousins.
Work to be done
That said, the rise of large language models, generative AI generally, and agentic AI does imply that I should update the agentive framework to take account of these things. Though a full accounting would be beyond the scope of this article, some first thoughts are below.
Since agentic AI allows natural language input, design should…
- Help users know what they can ask
- Help clarify any ambiguities in user requests
- Help users improve prompts to fit capabilities
Since agentic AI creates and executes a plan, design should…
- Expose the plan, in advance, during, and after execution.
- Allow a user to edit steps in the plan.
- Allow a user to establish or import forbidlists and requirelists that constrain execution.
- Allow a request to be re-run in full or from an adjusted step.
Since agentic AI takes time to execute, design should…
- Alert users with updates and when it is “complete.”
- Explain what complete means at this stage.
- Outline remaining steps to be done.
Since agentic AI iterates output over time, design should…
- Allow a user to review changes over time, and revert parts
- Allow users to make a one-time request into a persistent object.
- Allow users to subscribe to future updates.
Conclusio
As the fields of agentive technology and agentic AI continue to evolve and intersect, there’s an exciting opportunity for designers, developers, and researchers to create vastly more powerful and user-friendly systems. To stay ahead of the curve, professionals in both UX design and AI development should familiarize themselves with these concepts and understand their differences and synergies. (Sorry I had to write synergies.) Whether you’re designing user interfaces or building AI systems, consider how you can incorporate agentive principles and agentic AI capabilities to create more effective user-centric solutions. The future of human-AI interaction is being shaped now, and mastery of each will be important to setting them on the right track.
