I normally focus more on human language AI Agents as apposed to coding AI Agents. But I found this study interesting because it solves a common problem.
It has become clear now that AI Agents are not well suited for long running tasks. And a task is not a job…a job can be seen as a collection of tasks.
So AI Agents cannot be left unsupervised for long running tasks…
It became clear that AI Agents can exist in different paradigms based on the use-case and the environment they need to operate in.
Something I have also noticed experimenting with coding AI Agents, is that it is hard to give an explicit enough instruction for long running tasks.
Invariably there are gaps that are open for interpretation and the AI Agent revert to implicit assumptions.
This AI Agent framework addresses the problem of insufficient context and implicit decision-making by training the agent to step out of the way at those junctures — halting suggestions early to cede control back to the human user naturally, without interruption.
There has been a number of frameworks which calls on a human when confidence is low…where the human is defined as a tool together with other tools.
But, what sets this study apart is its passive, non-intrusive approach to human involvement.
Instead of actively querying the user (What do you mean by X?), the Empower method trains the LLM to detect ambiguity via entropy estimates and simply halts suggestions at those points, ceding control back to the human naturally.
Most AI training methods (like RLHF, where humans give direct feedback through ratings or questions) are expensive and time-consuming.
But Empower trains the AI using just old, existing data (like code written by people in the past).
]The paper calls out problems with AI's that ask too many questions, so it skips that entirely. Instead, the AI only handles the easy, routine stuff (like standard code snippets) and stops at tricky spots where things could go many ways.
This keeps the AI helpful without it trying to take over or trick people, which can happen in reward-based systems.
While the paper focuses on coding, the framework is generalisable…which I like…
Empowerment could apply to any sequential, text-based task, like writing reports, planning projects, or even assistive robotics described via language.
For instance, in a long-form writing assistant, the AI might handle transitions and formatting but pause at creative branches.
Or in web agents, it could automate routine navigation but halt at decision points like choosing search filters. The method's self-supervised nature — requiring only offline data — makes it scalable without constant human feedback.
Of course, limitations exist. It assumes humans will intervene naturally at stop points, and for non-text domains, you'd need to adapt the state representation.
Still, as AI Agents evolve, approaches like Empower could bridge the gap between short-term efficiency and long-term human-centric collaboration.
Chief Evangelist @ Kore.ai | I'm passionate about exploring the intersection of AI and language. Language Models, AI Agents, Agentic Apps, Dev Frameworks & Data-Driven Tools shaping tomorrow.
