In a few years, autonomous artificial-intelligence “agents” could be performing all sorts of tasks for us, and may replace entire white-collar job functions, such as generating sales leads or writing code.
Unlike basic chatbots we use today, these entities can venture out in the digital world and do things in our stead. They can log into accounts, communicate on our behalf via text and voice, write programs, and in theory do pretty much anything else we do with our computers. The implications of unleashing them on the world are likely to be small at first, but eventually they could realize the full promise — and peril — of AI.
Aaron Levie, chief executive of Box, the cloud-based file-sharing and content-management company, recently told me that the way AI can take over tasks that until recently could only be done by humans “is probably the biggest thing that’s ever happened” to his company.
It’s clear that AI agents are much more advanced in some domains than others, such as coding, where their precursors are already having a big impact on how that job is done.
In April, Cognition Labs, which makes Devin, a semiautonomous AI programmer and an early example of an AI agent, reached a $2 billion valuation, just six months after its founding, with $175 million from investors including Elad Gil, Khosla Ventures and Peter Thiel’s Founders Fund.
Dustin Moskovitz, who co-founded Facebook and is now CEO of Asana, a cloud-based work-management system for teams, says that fully autonomous AI agents are in their earliest stages, and too often fail at tasks.
Even so, his company is focused on enabling everyone — not just programmers — to create what may someday be viewed as the precursors to such agents, by letting people construct workflows that call on AIs to perform specific tasks.
Sales-development representatives are usually workers in their first sales job. They are tasked with finding, researching and making the initial outreach — via email, LinkedIn and the like — to a potential buyer. They are also early candidates for AI replacement.
Moskovitz’s Asana has already created routines that largely automate the role of a sales-development associate, but leave it to a human to review and approve any outreach messages before they’re sent.
London-based 11x’s autonomous AI agent, called Alice, takes this a step further, says CEO Hasan Sukkar. Alice works 24 hours a day with no human supervision to research potential clients and reach out to them all on its own. Currently, it drafts and sends the emails and other messages it creates.
Starting next month, 11x will begin testing a system that goes even further, and actually conducts the first voice phone call with a potential customer. Alice will engage them in a conversation to gather the information it needs — so long as the human opts in and agrees to field a call from a bot.
Alice uses a handful of different AI platforms to digest, generate and fact-check content. These include Perplexity for searching for information about potential sales targets, and summarizing it; ChatGPT for digesting that data; and Claude for crafting emails and other kinds of messages to be sent to them. As these platforms improve over time, so should Alice.
Along the way, Alice uses AIs to check its own work — for example, using ChatGPT to evaluate whether one of the other AIs in this pipeline is likely to have hallucinated information that’s incorrect. Using AI to check the work of other AIs is also the idea behind the two-person startup Maihem, launched in 2023.
All of this happens with no human supervision at all. Humans can check the work of Alice if they want, and when companies first start using Alice, their employees often do, says Sukkar.
Devin, from San Francisco-based Cognition Labs, took the AI world by storm in March, when a demo video was released showing its apparent ability to complete entire, complicated coding tasks on its own.
Later, some programmers took issue with that demo, asserting that it was edited in a way that was deceptive. One coder, a computer-vision expert, said that Devin had misunderstood what he had asked for when he posted a task to the freelance jobs board Upwork.
When I ask him about the controversy his company’s demo inspired, Cognition Labs CEO Scott Wu says that it’s “super early” for Devin, and that it will “often make mistakes.” Other programmers have documented how they’ve successfully used Devin to build entire apps from start to finish, he adds.
Recently, when I visited AI expert Ethan Mollick at Wharton, he showed me his own successful experiment with Devin, which had managed, with minimal prompting, to create an account on Reddit and start demanding payment for solving coding tasks.
Devin works by allowing its core AI agent to decide how best to use a number of tools, including large language models. It can use the web to do research, use an app, or build a website. Devin can also write and run code. And it has the ability to prompt its human overseer when it hits a problem it can’t solve — like a Captcha designed to keep bots like it from logging into certain services.
One validation of the approach behind Devin is that there are already a handful of open-source alternatives to it, notably OpenDevin and GPT Pilot, which can build everything from apps to web services all on their own. One test of their abilities shows they succeed only around one time in five, however.
For now, when an AI agent is asked to tackle something new, it’s best if a human is monitoring what it does, and redirecting it when it fails or goes astray — a concept known as “human in the loop.” This means these agents are less autonomous, and more like proactive, but still not fully independent, assistants.
While he’s bullish on its potential, Levie, of Box, says that for now AI is limited to doing individual, discrete tasks. Companies have stored vast amounts of information on Box’s systems, so AI is used primarily as a way to make all that data useful. For example, AI can be used to flag clauses in contracts that might be risky, or transcribe all the video a company has stored, and then allow people to ask questions of that corpus of text.
For companies offering a complete product for a task that’s already well understood, like 11x, that human involvement is already baked into the logic of the system, says Sukkar, the CEO. Defining the sequence of tasks that Alice is given to complete — in other words its business logic, or algorithm — was critical, and took a fair amount of experimentation, he adds.
What this means is that for the foreseeable future, many of us won’t be replaced with virtual employees, so much as we’ll be using them to perform parts of our jobs for us. In the process, we may also be training them to do even more.
Eventually, some virtual employees may not need us at all, and people who use these systems will essentially be their supervisors. Wu, of Cognition Labs, says one of his goals is to provide a team of AI-based coding assistants to anyone who needs them. Sukkar of 11x has similar goals, only the virtual employees his company will offer will work in sales.
If AI agents succeed at making humans more productive, it will lead to some jobs being eliminated. What’s unclear is how quickly that will happen.
“Five years from now, we could be in a totally different world, in which I send in a request to an AI, and the AI itself is doing all of the actions,” says Levie. “It will be just purely going off on its own, roaming around, kind of like a human does.”
