A new landscape of AI agents is emerging, defined not by a single breakthrough but by a diverse array of specialized tools automating tasks from corporate workflows to personal research. MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) has released its AI Agent Index, an ecosystem-wide analysis that categorizes and assesses 30 leading AI agent systems based on 1,350 data points. The study reveals a field dominated by enterprise applications, with significant variation in functionality and, critically, in the level of autonomy granted to these systems, raising important questions about control and risk.

The research identifies three primary categories of agents. The largest segment, comprising 13 of the 30 systems, is enterprise workflow agents. These are platforms designed to automate business tasks, with examples including Microsoft 365 Copilot, IBM watsonx Orchestrate, SAP Joule Studio, Salesforce Agentforce, and ServiceNow AI Agents. Close behind are chat applications with agentic tools, which account for 12 systems. This category includes general-purpose chat interfaces augmented with extensive tool access, such as Anthropic's Claude Code, OpenAI's ChatGPT Agent and Codex, as well as agents embedded in broader products like Manus AI. The third category is browser-based agents, a group of five systems whose primary interface is direct browser or computer interaction. These agents, which include Perplexity Comet, ChatGPT Atlas, and ByteDance's Agent TARS, are distinct from simple chat agents with web search due to their ability to perform background execution, trigger events, and conduct direct transactions.

When it comes to practical use, the top application for AI agents is research and information synthesis, a function present in 12 of the 30 agents covered. This capability spans both consumer-facing chat assistants and enterprise platforms. The second most common use case is workflow automation across business functions like HR, sales, support, and IT, enabled by 11 agents, most of which are enterprise products. A third significant function is GUI or browser automation for tasks like form-filling, ordering, and booking, found across seven of the models.

A key finding of the MIT index is the considerable variation in autonomy levels across different agent types. Chat-first assistants, which include systems like Anthropic Claude, Google Gemini, and the standard OpenAI ChatGPT, maintain the lowest levels of autonomy. These operate on a turn-based interaction model, executing a single set of actions before waiting for the next user prompt. On the higher end of the spectrum are browser-based agents, which offer users more limited opportunities for mid-execution intervention. For example, Perplexity's Comet agent performs tasks autonomously once a query is sent. This higher autonomy, coupled with their ability to operate in the background and execute transactions, presents elevated risks compared to more constrained chat interfaces.

The MIT CSAIL study serves as a crucial map of a rapidly evolving and often confusing landscape. By systematically categorizing agents by origin, interface, and function, it provides developers and businesses with a clearer understanding of the tools available. More importantly, it directly links an agent's design—particularly whether it is chat-based or browser-based—to its potential risk profile. The analysis suggests that as agents become more autonomous and capable of independent action in digital environments, the challenges of oversight, safety, and unintended consequences will become increasingly pressing. This foundational research lays the groundwork for more informed development and deployment decisions, highlighting that the future of agentic AI is not a monolith but a spectrum of tools with distinct capabilities and trade-offs.