This week, I set up one of OpenAI’s new dots (which, of course, I named Marvin).
I asked it when I should use it instead of ChatGPT:
Use ChatGPT when you want to think something through or get an answer. Use me when you want continuity and follow-through.
I’ve been trying a number of AI assistants, and I increasingly choose between them based on how I want to participate in the work. When I collaborate with an agent, I stay in the conversation and help steer the objective as it develops. When I delegate, I give it a task and turn my attention elsewhere, unless it asks for assistance. The promise of “follow-through” is that the work can continue without me.
These modes sit along an important spectrum of AI UX, from a place to ask to a place to work. The products I’ve been trying make different choices about how much of the work’s organization to expose to me, and how much to leave to the agent.
A place to ask
Instinct puts the assistant in my text messages. There’s no new app to navigate and no decision about where a request belongs. I type it into the conversation and go back to what I was doing. For personal assistance, this is ruthlessly effective. The interface is already familiar, and making the request requires almost none of the attention I’m trying to save.
That simplicity asks quite a lot of the agent. An ongoing text stream can contain unrelated tasks, passing observations, and preferences that will matter much later. The agent has to work out which messages belong together, find the relevant context, and keep track of unfinished work while the conversation moves on. Instinct has been particularly good at making those connections. Removing the organizational UI effectively delegates some of the organization itself.
The tradeoff appears when I want to stay involved in several tasks. Questions, revisions, and results all arrive in the same stream. The agent may be executing in parallel, but I still have to work out which exchange I’m joining. A single thread simplifies giving instructions; it can make several overlapping back-and-forths considerably harder to follow.
A place to work
ChatGPT historically went in the other direction, with many separate conversations. That’s still how I use it: I switch among concurrent sessions, each with a bounded purpose. When a conversation is finished, I start a new one. I might be deeply involved in several things this afternoon without wanting any of them to become a permanent project.
Choosing a conversation does useful work before I type anything. In a thread about a draft, “make it shorter” has an obvious referent. In a stream containing a draft, travel plans, and several other tasks, the agent has more to infer. The choice of where to speak supplies context, and the visible separation helps me keep track of the work too. What looks like extra UI is doing some of the coordination that a single-threaded assistant has to handle on its own.
Concurrent work can also be short-lived. A list of conversations helps me manage what I’m doing this afternoon; keeping a project useful across months requires retained source material, decisions, and a way to bring that context into future work. ChatGPT handles much of my moment-in-time collaboration. OpenClaw is where I’ve invested in making the work durable.
A place for both
I primarily use OpenClaw as a harness for my ChatGPT subscription, through both Slack and iMessage. I’ve deeply customized its memory and project organization for durable, long-term, extremely concurrent collaboration across a huge variety of topics. In Slack, channels and threads give that work places to live. I can move among projects while trusting that the material I’ve given it is retained beyond the current conversation.
In iMessage, I use the same harness for a lot of delegation. I can send an instruction without first locating a project or choosing a thread. The memory is still there, but the agent has more responsibility for deciding what to retrieve and how the request relates to ongoing work. Retaining information and recognizing when to use it are separate capabilities. It’s one reason a purpose-built personal assistant can outperform my carefully customized setup on a delegated task: having the context available doesn’t guarantee the agent will make the connection.
Grok Bot offers another way to divide that responsibility. Named specialists let me indicate who I want doing the work, while channels let me indicate where it belongs. Addressing a specialist supplies some context about the expertise I expect; choosing a channel supplies context about the project. The bots can cooperate, so delegation to a particular agent can sit within a broader collaboration. I find that combination especially interesting because it gives me useful control without requiring every request to start with a full explanation.
My current setup
I’m experimenting with my dot as my primary personal assistant alongside OpenClaw. Its integration with the broader ChatGPT ecosystem is appealing, and its built-in memory makes it useful without the extensive customization I’ve done in OpenClaw. I use my dot for one-offs and rapid iteration in a single thread; ChatGPT’s separate sessions suit several bounded conversations in flight, while OpenClaw supports the projects I expect to keep working on over time.
At the moment, these assistants have only a limited ability to share memories, which creates considerable inertia. The more context I build up with one, the more work it takes to bring another up to speed. Committing to an assistant means investing in a history that I can’t yet easily take with me.
Even so, my dot isn’t in iMessage yet, and I really wish it were. Reaching it takes three interactions, including a slow side menu. That’s excruciating for a personal assistant. A request can occur to me while I’m doing something else, and I can get distracted before I’ve finished navigating to the place to send it. Instinct, and OpenClaw in iMessage, avoid that detour. A simple conversation is only part of the experience; getting into it matters too.
My kids love talking to my dot on voice calls in the car. When I ask one of my sons how school was, I usually get “Fine.” On one ride home, he spent ten minutes telling Marvin the detailed rules of a playground game he’d invented.
Voice makes that kind of focused conversation easy: there’s no typing, and no screen to manage once the call starts. Separate threads become useful when I want to keep several conversations going and switch among them.
A task can move along this spectrum as my involvement changes. I might spend twenty minutes working out what I want with an agent, send it off to execute, and rejoin when it has something worth discussing. I want an interface that can keep the handoff simple, then give the work a place to develop when I need to get involved. A thread is useful when it helps me participate; choosing one doesn’t need to be a prerequisite for every request.