The spectrum of agentic interactions
Recently I've started asking myself: What I would fully delegate to an agent, what I would co-pilot, and what I will most likely keep doing myself. 🤔 I am deliberately not starting with my answers!
And then in general - How much of a workflow does the user actually want to give away?
I like to think about the answer as the spectrum of agentic interactions.
You and the agent work together. The work moves between both of you, step by step.
This is not a ranking 😄, and moving to the right is, in my opinion, not automatically progress. Sometimes doing it yourself is the product. Sometimes an agent beside you makes the experience better. Sometimes you just want the chore gone, and you want to fully offload the work.
The real interactions can land anywhere on the spectrum. The agent might help with one step and disappear for the next. The user might delegate the research, take over for the decision, and then hand execution back to the agent.
Let's explore it a bit 😁
On one end: I want to do it myself
There are plenty of things people do online because they enjoy the interaction itself.
Watching a video is not an inefficient way to obtain its summary. Reading an article is not a slow version of extracting five bullet points. Browsing a record shop, exploring a map before a holiday, looking through apartments, building a playlist, reading a discussion, or finding a gift can all be part of the experience.
In those moments, finishing faster might not be the goal at all.
The same can be true for workflows that are important, unfamiliar or simply personal.
An agent may still help around the edges - explain an unfamiliar term, find the right page, or answer a question - without taking over the main interaction. The user remains the driver because they want to see, understand or enjoy what is happening.
In the middle: let's do it together
The agent can collect information, suggest the next step, fill repetitive forms, compare options or notice something the user missed. The user can redirect it, inspect the result and take the control back at any moment.
Think about researching a topic while the agent organizes sources, explains complex terms and suggests further explorations.
I see co-browsing as a subset of this broader co-pilot interaction. In co-browsing, the user and agent share the same browser context: both can see the page, and both can act on it. But co-piloting can also happen in an editor, a shared document, a support conversation, a terminal or anywhere else the human and agent work together.
This middle can cover a lot of ground. At one point the agent might only suggest. Later it might execute ten steps while the user watches. For a critical action, it can stop and ask the user to review what is about to happen.
It is less like handing over a complete task and more like deciding, step by step, who should do the next part.
On the other end: please do it for me
At the far end, the user does not want to operate the workflow at all. They want to state the outcome and move on.
These are often chores: reschedule an appointment, collect several quotes, or update the same information across a few services.
The agent can come back for a approval: present the message before sending it, explain a cancellation fee, or ask which of two ambiguous options the user prefers.
At this end of the spectrum an interface becomes an implementation detail. The agent might use a website, an API, a tool or another agent - the user may not want to see any of it.
The same task moves for different people
I know it is tempting to place whole categories on this spectrum: entertainment on the left, work in the middle, chores on the right.
Here comes my favourite answer: "It depends..." 🤔
One person loves planning every detail of a holiday. Another only wants the tickets to appear in their calendar. One developer wants an agent to fix a bug and open a pull request. Another wants to inspect every log and form the hypothesis together.
The same person might choose differently when they are learning, tired, in a hurry, or dealing with a costly mistake.
Feasibility is personal too. A workflow can be technically possible to delegate and still feel unrealistic to a customer because they do not trust the result, cannot easily verify it, or have not built the habit yet. For someone else, the direct interface may be the thing that is difficult to use, and an agent can make the workflow possible in the first place.
So a use case does not have one permanent position. So - what does it depend on? I would say: the person, the moment, the risk, the available agent, and how easy it is to recover when something goes wrong.
A demo vs. real demand
There is one more thing I mentally separate when thinking about this spectrum: what works in a demo and what people will realistically use.
Demos often use flows familiar to almost everyone: book a flight, order food, buy a pair of shoes, schedule an appointment. That is a sensible choice. The audience understands the task immediately and can focus on what the agent is doing.
But the agentic interactions that are most important for me, personally - might look confusing during a presentation on stage 😅.
Co-debugging a production issue with an agent - imagine both of us investigating a log ingestion delay in us-east-1 region after a recent rollout, testing hypotheses and deciding where to look next.
That might be a realistic, high-value co-pilot workflow. Such demo might resonate with technical audience, but arguably requires way more context than "find me a winter jacket" kind of use-case.
Where does the business sit?
I do not think every business needs to attack every point on the spectrum immediately. But it helps to know which part of the customer journey you are serving, for whom, and why.
That means talking to customers before deciding that a workflow belongs to full delegation. It also means watching for different groups whose answer may be very different from your own.
The useful product boundary might not be one fixed mode. It might be a clean handover between modes: let the user start directly, invite an agent into the difficult part, and allow delegation when the user is ready 💡.
A little call to action
Ask yourself: How are you using agents today? Which interactions do you want to keep for yourself? Where do you want a co-pilot? Which chores would you happily delegate and only return for the approval?
Then ask people around you the same questions and I'll ask people around me. What is a realistic agent use case for them today?
Agents are already here, so this does not have to be only a prediction about the future. People are building habits now - and I suspect we might get some very interesting, very different perspectives 😀 - all of it can inform future investments.
At the same time let's think about what people really need, not only what they want. 😉
More soon. ✨