Olena Malik/Moment via Getty Images
Follow ZDNET: Add us as a preferred source on Google.
ZDNET’s key takeaways
- MIT released a list of top AI agents and their functionalities.
- The largest portion focuses on enterprise workflows.
- Research and information synthesis is the top use case.
Which autonomous or semi-autonomous agents are making the greatest impact on the world — and potentially your job — these days? Certain agents are hogging all the headlines lately, but there are a number of function-specific agents available to developers and users.
MIT’s CSAIL — the university lab dedicated to AI research — set out to identify and document the background and capabilities of these agents, with its findings detailed in its latest AI Agent Index. The researchers conducted an ecosystem-wide analysis of state-of-the-art AI agents across 1,350 data points.
Also: AI agents are fast, loose, and out of control, MIT study finds
What is the functionality and origin of leading agents? The researchers found that interfaces are the most abundant, followed closely by enterprise workflow platforms. They also uncovered risks shared across these agents, as explored by my ZDNET colleague Tiernan Ray.
Agents featured in the MIT index include the following:
- Anthropic Claude/Claude Code
- Google Gemini/Gemini CLI
- Manus AI
- OpenAI ChatGPT/ChatGPT Agent/Codex/AgentKit
- Perplexity
- Alibaba MobileAgent
- ByteDance Agent TARS
- Perplexity Comet
- IBM watsonx Orchestrate
- Microsoft 365 Copilot
- SAP Joule Studio
- Salesforce Agentforce
- ServiceNow AI Agents
Here are the three leading categories of agents identified by the researchers:
Enterprise workflow agents (13 systems of the 30 systems covered): These are platforms with agentic features for automating business tasks. Examples include Microsoft 365 Copilot and ServiceNow Agent.
Chat applications with agentic tools (12 systems): This category primarily includes chat interfaces with extensive tool access, according to the researchers. Examples include general-purpose coding agents such as Claude Code, as well as agents embedded in broader products such as Manus AI and ChatGPT Agent.
Also: I let Chrome’s AI agent shop, research, and email for me – here’s how it went
Browser-based agents (5 systems): These are agents whose primary interface is browser or computer use, with extensive browser/computer interaction tools. “They are distinct from chat agents with web search capabilities — ChatGPT web search, Claude web search — which primarily perform retrieval and summarization,” the researchers state. “Browser-based agents present higher risks through background execution, event triggers, and direct transactions.” Examples include Perplexity Comet, ChatGPT Atlas, ByteDance Agent TARS.
What are the most popular uses for AI agents?
Top use cases for AI agents, cutting across the above categories, include research and information synthesis, as seen in 12 of the 30 agents covered, spanning both consumer chat assistants and enterprise platforms. Right behind this functionality is workflow automation across business functions — such as HR, sales, support, and IT — enabled by 11 agents, primarily found in enterprise products. Agents focused on GUI or browser capabilities, used for tasks such as forms, ordering, and booking, are present across seven of the models.
Also: 6 reasons why autonomous enterprises are still more a vision than reality
Levels of autonomy vary considerably, the researchers found. Chat-first assistants maintain the lowest levels of autonomy. These are based on turn-based interactions, and include Anthropic Claude, Google Gemini, and OpenAI ChatGPT, which “executes a single set of actions and waits for the next user prompt.”
On the higher end of autonomy, browser agents offer more “limited opportunities for mid-execution intervention.” These include Perplexity’s Comet, which performs tasks autonomously once prompted. “Once a query is sent, users cannot easily intervene or steer the agent until it finishes.”
Enterprise platforms are split when it comes to agent autonomy. “During the design phase, users manually configure triggers, actions, and guardrails using visual canvases,” the researchers wrote. Others may offer AI assistance with this process. Once deployed, these agents often operate at higher levels of autonomy, “triggered by events like a new email or a database change, without any human involvement during the actual task execution.” Such agents include Glean, Google Gemini Enterprise, IBM watsonx, Microsoft 365 Copilot, n8n, and OpenAI AgentKit.
Also: Google Workspace now lets you create AI agents to automate your work – how to get started
A few offerings are considered developer/command-line-interface (CLI) agents that require explicit confirmation for sensitive operations such as file edits and command execution. Some agents offer “watch mode” for real-time oversight of critical actions, including ChatGPT Agent/Atlas, and Opera Neon.
Agent developers are concentrated in the US and China, with limited representation from other regions, the study also found.
