AI agents are considered the following big thing in AI, but there isn’t a exact definition of what they’re. So far, there isn’t a consensus on what exactly constitutes an AI agent.
In easy terms, an AI agent is best described as AI-powered software that performs a set of tasks for you which may previously have been done by a human customer support representative, HR representative, or IT help desk worker, although ultimately it might be any task. You ask it to do things, and it does them for you, sometimes across multiple systems and much beyond simply answering questions.
Seems easy enough, right? But it's complicated by a scarcity of clarity. There's no consensus even among the many tech giants. Google sees them as task-based assistants that modify depending on the duty: helping developers code; helping marketers create a color scheme; helping an IT skilled find an issue by querying log data.
At Asana, an agent can act like a further worker, taking good care of the tasks assigned to them like every good colleague. Sierra, a startup founded by former Salesforce co-CEO Bret Taylor and Google veteran Clay Bavor, sees agents as customer experience tools that help people perform actions that go far beyond the chatbots of yesterday and help solve more complex problems.
The lack of a typical definition actually leaves room for confusion about what exactly this stuff are imagined to do. But regardless of how they’re defined, the agents are supposed to help complete tasks in an automatic manner and with as little human interaction as possible.
Rudina Seseri, founder and managing partner of Glasswing Ventures, says it's still too early and that might be the explanation for the dearth of agreement. “There is not any single definition of an 'AI agent.' However, essentially the most common view is that an agent is an intelligent software system that perceives its environment, reason about it, makes decisions, and takes actions to autonomously achieve certain goals,” Seseri told TechCrunch.
She says they use a spread of AI technologies to realize this. “These systems integrate various AI/ML techniques comparable to natural language processing, machine learning and computer vision to operate in dynamic areas autonomously or in collaboration with other agents and human users.”
Aaron Levie, co-founder and CEO of Box, says that over time, as the facility of AI increases, AI agents will find a way to do quite a bit more for humans. And there are already dynamics in place that can drive this development.
“With AI agents, there are several components of a self-reinforcing flywheel that can help dramatically improve AI agent performance within the short and long run: GPU price/performance, model efficiency, model quality and intelligence, AI frameworks, and infrastructure improvements,” Levie wrote. is LinkedIn recently.
That's an optimistic view of the technology that assumes growth will occur in all of those areas, although that's not necessarily a given. MIT robotics pioneer Rodney Brooks identified in a recent TechCrunch interview that AI faces much harder problems than most other technologies and that it won't necessarily grow as quickly as, say, chips under Moore's Law.
“When a human sees an AI system perform a task, they immediately generalize that to similar things and estimate the competence of the AI ​​system; not only performance on that, but competence on that,” Brooks said in that interview. “And they're often very overoptimistic, and that's because they're using a model of how an individual performs on a task.”
The problem is that crossing systems is difficult. This is further complicated by the undeniable fact that some legacy systems lack basic API access. While we’re seeing the regular improvements Levie alluded to, getting software to access multiple systems and resolving issues that arise may prove tougher than many think.
If that's the case, everyone could also be overestimating what AI agents can do. David Cushman, research director at HFS Research, sees the present generation of bots more like Asana: assistants that help humans complete specific tasks to realize a user-defined strategic goal. The challenge is to assist a machine handle contingencies in a really automated way, and we're obviously a great distance from that.
“I feel that's the following step,” he said. “This is where AI works independently and effectively at scale. So that is where humans set the policies and guardrails and apply multiple technologies to take the human out of the loop – whereas GenAI has been about keeping the human within the loop,” he said. The key here is to let the AI ​​agent take over and apply true automation.
Jon Turow, partner at Madrona Ventures, says it will require the creation of an AI agent infrastructure, a technology stack specifically designed to create the agents (nonetheless you define them). In a recent blog post, Turow writes outlined examples of AI agents currently function within the wild and the way they’re built today.
Turow believes that the increasing proliferation of AI agents – and he also admits that the definition continues to be a bit elusive – requires a tech stack like every other technology. “All of this implies our industry must do work to construct an infrastructure that supports AI agents and the applications that depend on them,” he wrote within the article.
“Over time, considering will step by step improve, pioneering models will drive increasingly workflows, and developers will wish to deal with products and data – the things that differentiate them. They want the underlying platform to 'just work' and supply scalability, performance and reliability.”
Another point to take into accout here is that you most likely need multiple models, not only a single LLM, to get agents running, and that's comprehensible for those who consider those agents as a group of various tasks. “I don't think any single large language model, at the least not a publicly available, monolithic large language model, is able to handling agent-related tasks immediately. I don't think they’ll yet do the multi-step reasoning that may really excite me about an agent-related future. I feel we're getting closer, but we're not there yet,” said Fred Havemeyer, head of US AI and software research at Macquarie US Equity Research.
“I feel essentially the most effective agents will probably be multiple collections of multiple different models with a routing layer that sends requests or prompts to essentially the most effective agent and model. And I feel it could be something like an interesting (automated) supervisor, delegator role.”
Ultimately, Havemeyer says the industry is working toward the goal of agents operating independently of each other. “When I feel concerning the way forward for agents, I need to see agents which are truly autonomous and might pursue abstract goals after which think through all the person steps in between completely independently, and that's what I hope,” he told TechCrunch.
In reality, we’re still in a transition phase in the case of these agents, and we don't know when we’ll reach this end state described by Havemeyer. While what we’ve seen thus far is clearly a promising step in the correct direction, we still need some progress and breakthroughs to make AI agents work as they’re intended today. And it is necessary to know that we are usually not there yet.
