AI Agents Are Here. What Now? Back to Articles a]:hidden"> AI Agents Are Here. What Now? Published January 13, 2025 Update on GitHub Upvote 88 +82 Margaret Mitchell meg Follow Avijit Ghosh evijit Follow Sasha Luccioni sasha Follow Giada Pistilli giadap Follow This article is also available in Chinese 简体中文. Introduction The sudden, rapid advancement of LLM capabilities – such as writing fluent sentences and achieving increasingly high scores on benchmarks – has led AI developers and businesses alike to look towards what comes next: What game-changing technology is just on the horizon? One technology very recently taking off is “AI agents”, systems that can take actions in the digital world aligned with a deployer’s goals. Most of today’s AI agents are built by incorporating large language models (LLMs) into larger systems that can perform multiple functions. A fundamental idea underlying this new wave of technology is that computer programs no longer need to function as human-controlled tools, confined to specialized tasks: They can now combine multiple tasks without human input. This transition marks a fundamental shift to systems capable of creating context-specific plans in non-deterministic environments. Many modern AI agents do not merely perform pre-defined actions, but are designed to analyze novel situations, develop relevant goals, and take previously undefined actions to achieve objectives. In this piece, we briefly overview what AI agents are and detail the ethical values at play, documenting tradeoffs in AI agent benefits and risks. We then suggest paths forward to bring about a future where AI agents are as beneficial as possible for society. For an introduction to the technical aspects of agents, please see our recent developer blogpost. For an introduction to agents written before modern generative AI (that is largely still applicable today) please see Wooldridge and Jennings, 1995. Our analysis reveals that risks to people increase with a system’s level of autonomy: The more control a user cedes, the more risks arise from the system. Particularly concerning are risks for the safety of individuals that arise from the same benefits that motivate AI agent development, such as freeing developers from having to predict all actions a system may take. Further compounding the issue, some safety harms open the door for other types of harm – such as harms of privacy and security – and inappropriate trust in unsafe systems enables a snowball effect of yet further harms. As such, we recommend that fully autonomous AI agents are not developed. For example, AI agents that can write and execute their own code, beyond constrained code options controlled by the developer, will be endowed with the ability to override all human control. In contrast, semi-autonomous AI agents may have benefits that outweigh risks, depending on the level of autonomy, the tasks available to the system, and the nature of individuals’ control over it. We now turn to these topics in-depth. What is an AI agent? Overview There is no clear consensus on what an “AI agent” is, but a commonality across recently introduced AI agents is that they are “agentic”, that is, they act with some level of autonomy: given the specification of a goal, they can decompose it into subtasks and execute each without direct human intervention. For example, an ideal AI agent could respond to a high-level request such as “help me write better blogposts” by independently breaking this task down into retrieving writing on the web that is similar to your previous blog topics; creating documents with outlines for new blog posts; and providing initial writing within each. Recent work on AI agents has made possible software with a broader range of functionality and more flexibility in how it can be used than in the past, with recent systems deployed for everything from organizing meetings (example1, example2, example3, example4) to creating personalized social media posts (example), without explicit instructions on how to do so. All recently introduced AI agents we’ve surveyed for this newsletter are built on machine learning models, and most specifically use large language models (LLMs) to drive their actions, which is a new, novel approach for computer software. Aside from being built on machine learning, today’s AI agents share similarities with those in the past, and in some cases realize previous theoretical ideas of what agents might be like: acting with autonomy, demonstrating (perceived) social ability, and appropriately balancing reactive and proactive actions. These characteristics have gradations: Different AI agents have different levels of capabilities, and may work in isolation or in concert with other agents towards a goal. As such, AI agents may be said to be more or less autonomous (or agentic), and the extent to which something is an agent may be viewed on a continuous spectrum. This fluid notion of AI agent has led to recent confusions and misunderstandings about what AI agents are, which we hope to bring some clarity to here. A table detailing the varying levels of AI agent is provided below. Agentic Level Description Who's in Control What that's Called Example Code ☆☆☆☆ Model has no impact on program flow 👤 The developer controls all possible functions a system can do and when they are done. Simple processor print_llm_output(llm_response) ★☆☆☆ Model determines basic control flow 👤 The developer controls all possible functions a system can do; the system controls when to do each. Router if llm_decision(): path_a() else: path_b() ★★☆☆ Model determines how function is executed 👤 💻 The developer controls all possible functions a system can do and when they are done; the system controls how they are done. Tool call run_function(llm_chosen_tool, llm_chosen_args) ★★★☆ Model controls iteration and program continuation 💻 👤 The developer controls high-level functions a system can do; the system controls which to do, when, and how. Multi-step agent while llm_should_continue(): execute_next_step() ★★★★ Model writes and executes new code 💻 The developer defines high-level functions a system can do; the system controls all possible functions and when they are done. Fully autonomous agent create_and_run_code(user_request) Table 1. One example of how systems using machine-learned models, such as LLMs, can be more or less agentic. Systems can also be combined in "multiagent systems," where one agent workflow triggers another, or multiple agents work collectively toward a goal. Adapted from smolagent blog post, with changes tailored for this blog post. From an ethics perspective, it is also useful to understand the continuum of autonomy in terms of how control is ceded from people and given to machines. The more autonomous the system, the more we cede human control. Throughout this piece, we use some anthropomorphising language to describe AI agents, consistent with the language that is currently used to describe them. As was also noted in historic scholarship, describing AI agents using mentalistic language ordinarily applied to humans – such as having knowledge, beliefs, and intentions – can be an issue for appropriately informing users about system abilities. For better or worse, such language serves as an abstraction tool to gloss over more precise details of the technology. Understanding this is critical when grappling with the implications of what these systems are and the role they may play in peoples’ lives: The use of mentalistic language describing AI agents does not entail that these systems have a mind. The Spectra of AI Agents AI agents vary on a number of interrelated dimensions: Autonomy: Recent “agents” can take at least one step without user input. The term “agent” is currently used to describe everything from single-step prompt-and-response systems (citation) to multi-step customer support systems (example). Proactivity: Related to autonomy is proactivity, which refers to the amount of goal-directed behavior that a system can take without a user directly specifying the goal (citation). An example of a particularly “proactive” AI agent is a system that monitors your refrigerator to determine what food you are running out of, and then purchases what you need for you, without your knowledge. Smart thermostats are proactive AI agents that are being increasingly adopted in peoples’ homes, automatically adjusting temperature based on changes in the environment and patterns that they learn about their users’ behavior (example). Personification: An AI agent may be designed to be more or less like a specific person or group of people. Recent work in this area (example1, example2, example3) has focused on designing systems after the Big Five personality traits – Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism as a “psychological framework” (citation) for AI. At the end of this spectrum would be “digital twins” (example non-agentic digital twin). There are currently not agentic digital twins that we are aware of. Why creating agentic digital twins is particularly problematic has recently been discussed by the ethics group at Salesforce, among others (example). Personalization: AI agents may use language or perform actions that are aligned to a user’s individual needs, for example, to make investment recommendations based on current market patterns and investments a user has made in the past. Tooling: AI agents also have varying amounts of additional resources and tools they have access to. For example, the initial wave of AI agents accessed search engines to answer queries, and further tooling has since been added to allow them to manipulate other tech products, like documents and spreadsheets (example1, example2). Versatility: Related to above is how diverse the actions that an agent can take are. This is a function of: Domain specificity: How many different domains an agent can operate in. For example, just email, versus email alongside online calendars and documents. Task specificity: How many different types of tasks the agent may perform. For example, scheduling a meeting by creating a calendar invite in participants’ calendars (example), versus additionally sending reminder emails about the meeting and providing a summary of what was said to all participants when it’s over (example). Modality specificity: How many different modalities that an agent can operate in – text, speech, video, images, forms, code. Some of the most recent AI agents are created to be highly multimodal (example), and we predict that AI agent development will continue to increase multimodal functionality. Software specificity: How many different types of software the agent can interact with, and at what level of depth. Adaptibility: Similar to versatility is the extent to which a system can update its action sequences based on new information or changes in context. This is also described as being “dynamic” and “context-aware”. Action surfaces: The places where an agent can do things. Traditional chatbots are limited to a chat interface; chat-based agents may additionally be able to surf the web and access spreadsheets and documents (example), and may even be able to do such tasks via controlling items on your computer’s graphical interface, such as by moving around the mouse (example1, example2, example2). There have also been physical applications, such as early-stage agents embodied in robots (example). Request formats: A common theme across AI agents is that a user should be able to input a request for a task to be completed, without specifying fine-grained details on how to achieve it. This can be realized with low-code solutions (example), with human language in text, or with voiced human language (example). AI agents whose requests can be provided in human …