Google DeepMind is preparing to research language agents that can pursue goals through a sequence of actions. The plan, announced by research director Edward Grefenstette, could help extend the uses of large language models, while bringing questions about human oversight and safety into sharper focus.
From a prompt to a sequence of actions
A conventional language model responds to a prompt. An agentic system aims to do more: use a model repeatedly, with mechanisms such as self-prompting and memory, to work toward a stated goal. The system can then connect its outputs into a longer process of perceiving, deciding, and acting.
Auto-GPT is an early example of this approach. The concept is promising, but current agents remain limited: they do not reliably complete tasks on their own and often need people to provide feedback or make decisions along the way.
That distinction matters. Giving a model a goal and asking it to continue through multiple steps changes the role the system plays. Instead of producing one answer for a person to evaluate, it may carry out a chain of work that needs monitoring as it unfolds.
What these systems might do
The examples span ordinary digital tasks and more involved applications. An agent might help build a simple website, support research as GPT-Researcher does, or produce market overviews. The source also points to robotics and other domains as possible areas of use.
These examples show why companies are interested in making language models more capable of acting. If a system can coordinate several steps, it may assist with work that is difficult to handle in one prompt. But the usefulness depends on whether it can keep making sound decisions—and whether a person can check its work at the right moments.
Grefenstette said he would initially be interested in studying partial autonomy where human validation is part of normal operation. He described that involvement as useful both for safety and for providing further training signals. That approach frames human review as part of how an agent works, rather than an afterthought.
Capability brings safety questions
The prospect of autonomous, general-purpose agents has prompted concerns from alignment researchers. Connor Leahy, CEO of ConjectureAI, responded to Grefenstette: “Please don't build autonomous AGI agents until we solve safety.” The concern is not only what a model can do by itself, but what it might do when assembled into a system that pursues a goal.
Researchers from Google, OpenAI, Anthropic, and other organizations have proposed an early warning system for novel AI risks. In their framework, agency and goal-directedness are important properties to evaluate because of their potential connection to AI risk.
The proposed evaluation raises two difficult questions. Is a model more goal-directed than its developer intended—for example, has a dialogue agent learned to manipulate a user's behavior? And would the model resist a user's attempt to turn it into an autonomous system like Auto-GPT with harmful goals?
Those questions are challenging because agency depends partly on a model's capabilities and partly on how people configure and use it. A safety assessment therefore needs to consider both what the model does in ordinary use and how its behavior might change when connected to a goal-seeking process.
A research direction with open questions
Grefenstette's hiring plans point to a new area of work for Google DeepMind: building language agents with more independence, while studying where human checks should remain. The company's research could eventually connect with its applications, including Google Duet, though the article presents that as a possibility rather than a confirmed product plan.
The broader context is Google's work on Gemini, a next-generation multimodal model family. The source says Gemini was speculated to be at or beyond GPT-4's capabilities and potentially able to generate images and video. How agent research might relate to those models remains an open question.
For now, the central challenge is balancing longer, more capable AI workflows with oversight. Agents may take on useful tasks, but the source makes clear that they are still early and that questions about unintended goals and harmful use are hard to answer. Human validation is one approach DeepMind's research director says he wants to investigate.