AI has made it much faster to build, research, and write, but we still spend a lot of time managing the chats. I think the next big shift is an assistant that manages more of that work, including the follow-up prompts. You describe what you want and how it should work, and it keeps the other agents moving toward that result. I've been exploring this with Dottie, my personal OpenAI dot, and I think it is a bigger change than it first appears.
Outline
We Are Still Managing the Chats
You open Claude Code, Cursor, or another coding agent and ask it to build a feature. It makes changes and comes back with a result. You test it, find something that is off, and explain what needs to change.
Maybe you open another chat to research an approach or review the code. Then you bring those findings back to the coding agent. It takes another pass, and you test again.
The agent is doing a lot of the implementation, but you are still connecting the pieces. You decide which chat needs which information, read the results, and send the next instruction. And... that can become a job of its own.
In my previous post on AI Agents, I covered how agents use MCPs, Skills, Tools, and subagents to work toward an objective. Now I want to look at who manages the ongoing conversations around that work.
A Real Example With Dottie
I've been working with Dottie on a visual Bible explorer. I wanted to explore topics at several levels, similar to a pillar and spoke model in marketing, with related passages under each topic. I explained the idea as I was thinking through it:
My original explanation and Dottie's response in the OpenAI app.
I explain the experience I want, and Dottie suggests expandable topic groups, breadcrumbs, and a passage list. She also points out that promising "all verses" would be too strong, since AI grouping can miss connections.
Later in the same project, I can open the coding task and see a follow-up labeled Sent via Dottie:
A later follow-up from Dottie to the agent working on the prototype.
Dottie passes along my request, then turns it into instructions for the coding agent: expandable topics and subtopics, breadcrumbs, and a passage list coordinated with the graph. She also asks for topic labels grounded in the passages and carries forward the caution about promising "all verses." The agent gets both the idea and the details needed for the next iteration.
Dottie is writing the next instruction in the task's conversation. I can stay focused on the product while she handles more of the prompting needed to move it forward.
The Important Part Is the Follow-Up
The primary assistant needs to read what came back, compare it with the objective, and decide what to ask next. A coding agent might have built the feature but missed an edge case. A researcher might have answered most of the question but left a gap. The next prompt depends on those results.
You can't prepare every useful follow-up before the work starts. Someone has to notice what is missing and carry that feedback into the next attempt.
OpenAI's tasks and memory guide describes dots creating separate threads, checking results, and sending follow-up instructions to tasks they created. Anthropic has described a related lead-agent approach in its multi-agent research system.
Agents coordinating other agents is an established idea. What I think is changing is how much of that coordination becomes part of the assistant we talk to every day.
Do We Still Need to Be Prompt Engineers?
I think the burden on the person using AI will drop quite a bit. You should be able to explain a goal naturally, work through the decisions with your assistant, and let it prepare the instructions for the agents doing the work.
An assistant can write those prompts for us, but we still need to decide what a good outcome looks like. This connects to the north star I wrote about before. For our coffee shop example, the goal might be that the manager can see which beans need ordering in a few seconds. The assistant can help turn that goal into implementation tasks, but we still need to decide whether the result solves the problem.
Examples, preferences, and constraints become valuable direction. Use our existing design. Keep it within this budget. Show me a working version before deploying. Those details give the assistant something to check against when another agent says the work is done.
The messages it sends may look more like task notes than our original conversation. I explored that wording, and how to try it yourself, in Prompting AI vs. Talking to Humans.
Why I Think This Is a Bigger Change
If every agent needs you to read its output and type the next instruction, your attention becomes the bottleneck. Running more agents can simply give you more chats to manage.
An assistant that takes on those intermediate steps changes how much work you can delegate. Someone who understands a business problem could describe it without also learning how to manage coding terminals, research chats, and review prompts.
I expect this to matter beyond software too. A proposal might need research, calculations, writing, and review. A campaign might need copy, visuals, and a landing page. When one piece changes, the assistant can carry that decision into the other pieces.
Reliability will matter more than the number of agents running. Can the assistant inspect the actual result? Does it recognize when something is incomplete? Can it tell when it needs your judgment, and can you see and redirect what it is doing?
Delegation adds cost and complexity, so a small task may still be best handled by one agent. Choosing when to delegate is part of the job.
Wrapping Up
I want to explain a change to Dottie and have her work through the revisions with the coding agent, then show me what changed.
How much of that coordination would you hand over to an assistant? Where would you still want to be involved? Drop your thoughts in the comments :)
Comments
Login to Add comments.