No one knows what the hell an AI agent is

Silicon Valley is bullish on AI brokers. OpenAI CEO Sam Altman mentioned brokers will “be part of the workforce” this yr. Microsoft CEO Satya Nadella predicted that brokers will exchange sure data work. Salesforce CEO Marc Benioff mentioned that Salesforce’s aim is to be “the primary supplier of digital labor on the earth” by way of the corporate’s numerous “agentic” companies.

However nobody can appear to agree on what an AI agent is, precisely.

In the previous couple of years, the tech business has boldly proclaimed that AI “brokers” — the newest buzzword — are going to vary all the things. In the identical manner that AI chatbots like OpenAI’s ChatGPT gave us new methods to floor data, brokers will basically change how we strategy work, declare CEOs like Altman and Nadella.

That could be true. However it additionally relies on how one defines “brokers,” which isn’t any simple activity. Very like different AI-related jargon (e.g. “multimodal,” “AGI,” and “AI” itself), the phrases “agent” and “agentic” have gotten diluted to the purpose of meaninglessness.

That threatens to go away OpenAI, Microsoft, Salesforce, Amazon, Google, and the numerous different firms constructing complete product lineups round brokers in a clumsy place. An agent from Amazon isn’t the identical as an agent from Google or every other vendor, and that’s resulting in confusion — and buyer frustration.

Ryan Salva, senior director of product at Google and an ex-GitHub Copilot chief, mentioned he’s come to “hate” the phrase “brokers.”

“I believe that our business overuses the time period ‘agent’ to the purpose the place it’s nearly nonsensical,” Salva instructed Trendster in an interview. “[It is] one among my pet peeves.”

The agent definition dilemma isn’t new. In a bit final yr, former Trendster reporter Ron Miller requested: What’s an AI agent? The issue he recognized is that almost each firm constructing brokers approaches the tech otherwise.

It’s an issue that’s worsened just lately.

This week, OpenAI printed a weblog submit that outlined brokers as “automated programs that may independently accomplish duties on behalf of customers.” But in the identical week, the corporate launched developer documentation that outlined brokers as “LLMs outfitted with directions and instruments.”

Leher Pathak, OpenAI’s API product advertising and marketing lead, later mentioned in a submit on X that she understood the phrases “assistants” and “brokers” to be interchangeable — additional muddying the waters.

In the meantime, Microsoft’s blogs attempt to distinguish between brokers and AI assistants. The previous, which Microsoft calls the “new apps” for an “AI-powered world,” might be tailor-made to have a selected experience, whereas assistants merely assist with normal duties, like drafting emails.

AI lab Anthropic addresses the hodgepodge of agent definitions just a little extra instantly. In a weblog submit, Anthropic says that brokers “might be outlined in a number of methods,” together with each “totally autonomous programs that function independently over prolonged intervals” and “prescriptive implementations that observe predefined workflows.”

Salesforce has what’s maybe essentially the most wide-ranging definition of AI “agent.” In accordance with the software program big, brokers are “a kind of […] system that may perceive and reply to buyer inquiries with out human intervention.” The corporate’s web site lists six totally different classes, starting from “easy reflex brokers” to “utility-based brokers.”

So why the chaos?

Properly, brokers — like AI — are a nebulous factor, and so they’re continuously evolving. OpenAI, Google, and Perplexity have simply began transport what they contemplate to be their first brokers — OpenAI’s Operator, Google’s Challenge Mariner, and Perplexity’s buying agent — and their capabilities are everywhere in the map.

Wealthy Villars, GVP of worldwide analysis at IDC, famous that tech firms “have an extended historical past” of not rigidly adhering to technical definitions.

“They care extra about what they’re attempting to perform” on a technical stage, Villars instructed Trendster, “particularly in fast-evolving markets.”

However advertising and marketing can be accountable largely, in response to Andrew Ng, the founding father of AI studying platform DeepLearning.ai.

“The ideas of AI ‘brokers’ and ‘agentic’ workflows used to have a technical that means,” Ng mentioned in a current interview, “however a few yr in the past, entrepreneurs and some massive firms received a maintain of them.”

The dearth of a unified definition for brokers is each a chance and a problem, Jim Rowan, head of AI for Deloitte, says. On the one hand, the anomaly permits for flexibility, letting firms customise brokers to their wants. On the opposite, it could — and arguably already has — result in “misaligned expectations” and difficulties in measuring the worth and ROI from agentic initiatives.

“With no standardized definition, a minimum of inside a company, it turns into difficult to benchmark efficiency and guarantee constant outcomes,” Rowan mentioned. “This may end up in diversified interpretations of what AI brokers ought to ship, probably complicating mission targets and outcomes. In the end, whereas the pliability can drive inventive options, a extra standardized understanding would assist enterprises higher navigate the AI agent panorama and maximize their investments.”

Sadly, if the unraveling of the time period “AI” is any indication, it appears unlikely the business will coalesce round one definition of “agent” anytime quickly — if ever.