We’re at an interesting moment right now when it comes to artificial intelligence. For a couple of years, or really, ever since the first stirrings of publicly available “retail” AI tech, we’ve been scrutinizing what you might call the “first wave” of AI: LLMs that can approximate human cognition, but only in sort of a “Google search” way, providing, sorting, and generally dealing with information, in some ways an extension of an I/O model, albeit an amazingly human-like one.

Now, though, we’re moving toward something new: autonomous task engines. As people continue to ponder: can we make AI entities sentient?, others are suggesting that, sentient or not, they can handle routine tasks for us, or even longer-frame, multi-step tasks.

In that context, there’s a lot of thinking about what this will look like. Take an interview with Jenny Lay-Flurrie Trusted Technology Group set up by Microsoft in early 2025, in which Lay-Flurrie talks about advocacy for the blind and other aspects of equitable AI development.

“How do we make sure that we build it right?” she said. “And how can we make sure that it stays right? Society is not always the most inclusive place, so there are instances where we have to insert data to train it.”

Here’s more on evaluating the advent of “personal agents,” task-assisters that will make our lives easier, with a group of experts in a panel at the Imagination in Action event last month. (Disclaimer: April’s IIA event is an annual conference that I help to facilitate.)

Maria Gorskikh, of our own MIT community, mentioned OpenClaw, and also alternatives that people are working on now. Gorskikh works on project NANDA, something I have written about a lot in the past.

“What happened that was really important this year was OpenClaw,” she said. “OpenClaw was specifically important because it kind of kicked off the mass adoption of personal AI agents and it went viral: and now non- tech people and tech people together are excited about the world of agents and the agentic web.”

Panelist Jordan Tian, co-founder of ZeroClawLabs, talked about shrinking the hardware footprint with Raspberry Pi and other components.

“The idea was, we wanted to go all the way down and make something that could be small and deployed on any hardware,” Tian said, touting the value of AI edge computing. “At the time, everyone's looking at sandboxing and getting their own Mac Mini, and we were, like, what if you could make it smaller?”

“We are really looking into making a very specific use case, that is for the people who do not have the agency to give consent, because personal agents need consent,” said panelist Uzma Farheen, Director of Keep AI Safe. “There are two aspects to it. One is the infrastructure and the technical capabilities, and the next is the usage. For that, we are creating the platform to actually create the agents who can talk to each other, and simulate human interactions, to see where the agents and the models break.”

Calling for frameworks to handle these test cases, she mentioned the urgency on guardrail development.

“When innovation happens, safety kind of goes on the back burner,” she said.

Panelist Greg Raiz, general partner at Founder’s Edge, went over some thoughts on user design.

“I think we're at a really interesting inflection point in history,” he said. “I run a small venture firm. I think a lot about user experience. And when we look back in history, the biggest shifts are just changes in input and output.”

He noted the nascent status of personal AI.

“We can use our voice, we can delegate,” he said, “and again, we're in the early innings of building trust.”

Panelist Niresh Agarwal agreed.

“There is lot of anxiety as well as curiosity in the consumer mind,” he said. “AI agents are coming, but it's hard to understand what it will be like.”

Some Challenges and Limitations

Speaking about obstacles to an expedited design process and the invasion of AI agents, Gorskikh mentioned two costs that will factor in: one, the cost of hosting, and two, the cost of inference.

“I think we really need to solve these two barriers in order for everyone to have an agent that is cheap, reliable, secure, and accessible,” she said.

Later in the conversation, moderator Gunjan Sinha asked the panel how they think about the end goal in assistive AI agents.

“What does success look like here?” he asked. “Agents that are really winning for their consumers: is it going to these use case specifics that people value, or is it going to be more horizontal, applying in different arenas as orchestrators?”

Gorskikh theorized on AI agents that are specialized to a user’s profession. Tian discussed the incremental development of personal technology, with major landmarks along the way.

“You can give it objectives, and it should be able to just do these things,” he said. “The scope will grow as apps are tailored to allow them to do these types of things. I think that it'll go: personal AI assistant, next generation, and then it can be even more autonomous and keep doing things like this for you, that you would normally have to work through your phone to do.”

Raiz mentioned the power of AI agents to curate information for users, centralizing their focus in ways that reverse some of the earlier effects of AI on our brains, citing a paper called “Attention is All You Need” that came out a few years ago.

“AI has actually split our attention in multitudes of directions,” he said. “I think agentic tools start to be a success where they're giving our attention back, because they are surfacing the things that are actually important to us.”

As an additional orienting strategy, Agarwal suggested people should look at agents that are popular now, and extrapolate.

Sinha had a take on Gorskikh’s idea of personal specialization, and on Raiz’s notes on attention:

“Is it possible that the agents are here to help fix our ADHD that they created in the first place for us?” he asked.

I thought this was an interesting thread to pull, as so many of us intuitively feel that new technologies have splintered our own human ability to focus and center in on any given stimulus. Then we imagine that these agents will do some sort of tasking that will help solve this problem.

“The first personal ‘chief of staff’ agent that I built was exactly like this,” Raiz said. “The prompt was like, ‘Help me stay on task and on focus, because there's so many things distracting.’ Ais are really good at that, because they can synthesize information from a lot of disparate sources, and then create containers of priority, which kind of keep you more on track.”

Noting the history of agentic AI, Agarwal called for major advances, soon.

“Using the example of coding agents, they didn't make incremental improvement of software developers’ lives,” he said. “They made their lives 10 times, 100 times better. That's why it's being used. We need to figure out those use cases, immediate use cases. We need to figure out a way that any person and every person is able to say, ‘yes, I cannot live without it.’”

Detailing some of her own research, Gorskikh talked about how AI agents can help human users deal with “the boring stuff,” but said she does not yet trust OpenClaw to run an inbox. Tian mentioned the ability of agents to help users with browsing preference, automating feeds.

A conversation then ensued where panelists started breaking down the distribution of tasks and data between open source models, and closed systems that will protect a user’s privacy. Some brought up distrust or skepticism on an important question: are the agents built for augmentation, or engagement? Agarwal pointed out the common consumer mindset: not whether systems are secure, or focused on the right use cases, but whether they are offered for free, or not.

All of the above makes you think, of a world without traditional boundaries. How we navigate that world will be different. We need to get started.