Always-On Agents Tested for Task Discovery

As autonomous AI assistants move from answering explicit prompts toward acting on a user's behalf, a new research project asks a basic question: who notices the work? AutoBuddy: Evaluating Proactive, Always-On Agents on Task Discovery and Execution in a Simulated Workplace is the focus of a source report from the Grafy Research Lab. According to the source, the study examines proactive task discovery and execution, in which an agent must infer a user's pending tasks from their workspace and finish those tasks inside the tools where the work actually happens.

What did the study examine?

The source describes AutoBuddy as an evaluation of proactive, always-on agents. According to the material, the broader shift under study is that LLM agents are becoming more autonomous and are expected to serve as always-on assistants. The source says that existing benchmarks often supply the task directly. In contrast, the knowledge work setting examined involves requests that already sit in a user's workspace, meaning the agent has to recognise the obligation before acting.

The feed summary frames the motivation directly. It notes that as LLM agents become more autonomous, they are increasingly expected to run as always-on assistants on a user's behalf, spotting tasks that need action and then finishing them without a prompt. The study is therefore positioned at the boundary between autonomous agents and everyday task management.

The source reports that the study was published by the Grafy Research Lab and lists authors including Shiqi Chen, Daniel Madrigal Diaz, Wei-Ning Chen, Gbola Afonja, Fangkai Yang, Rujia Wang, Chetan Bansal, Victor Rühle, Saravan Rajmohan and Dongge Han, with subject category Artificial Intelligence. The available source material does not provide a complete methodology, so the public description should not be read as a full technical report.

Why task discovery is central

The source frames the core challenge. An always-on agent cannot wait for a prompt if it is meant to operate continuously. It must find tasks, judge what is relevant, and then carry them through completion. The source describes two challenges for the agent, though the available summary does not spell them out in full. The point is that discovery precedes execution in this model.

For AI assistants, this changes the division of labour. The model may still be capable of performing a task, but the system around it must decide when the task exists. The source suggests that in knowledge work, the requests are not delivered as clean commands. They are present in the workspace, and the agent must sort them out.

How the research approached proactive execution

According to the source, the agent is expected to complete tasks in the systems where the work lives. That detail matters because it moves the agent from conversation into action within existing work environments. The source does not provide a full technical account of how AutoBuddy was built or tested, so the public description remains limited.

The source material also frames the study as an evaluation. It does not present the always-on behaviour as already proven or broadly deployed. Instead, it places the challenge at the centre of the research question. The source lists the subject area as Artificial Intelligence, but the provided material does not include results, metrics or evaluation details.

What the source means for autonomous assistants

The source does not claim that always-on agents have solved proactive task discovery. Instead, it presents the scenario as an evaluation. For creators, developers and researchers, the practical lesson is that future assistants may need workspace awareness, memory and action controls. The source frames this as a challenge for the agent, not only for the user.

The report is likely aimed at people building AI products who need to decide how an agent should handle unasked obligations. It raises questions about when an agent should act, how it should prioritise, and how it should confirm work. The source does not answer all of those questions, and the available summary leaves much of the technical detail unclear. For researchers, this suggests a possible direction for agent benchmarks. For developers, it suggests that action space design may be as important as model choice.

How this fits with Grafy

For readers who want to examine the source directly, Grafy Reader can be used to read, search and question PDFs and papers with AI. Grafy Chat can support discussion of the paper's ideas. The provided source material does not describe a direct consumer feature from this study, so Grafy News treats the work as research context rather than a product launch.

Frequently asked questions

Does AutoBuddy teach a smaller agent, or is it taught by one?

The source does not say AutoBuddy teaches a smaller agent. According to the available material, AutoBuddy is the subject being evaluated for proactive task discovery and execution. The source does not establish whether a smaller agent is taught by AutoBuddy, or whether AutoBuddy learned from one. Readers should treat that question as unanswered by this source.

Was proactive task discovery included in a release that added always-on behavior?

The source does not frame proactive task discovery as a release feature. According to the material, AutoBuddy is a study of always-on agents. The available source does not describe a release, version, or timing that included such a feature. So the claim should not be read as a product release announcement.

Does the simulated workplace study prove always-on agents can complete tasks without being asked?

The source reports that the study examines proactive task discovery and execution, where an agent must infer pending tasks and complete them without being asked. It presents this as a challenge. The available source does not claim the study proves the agents can do so reliably. It describes evaluation of that setting, not a settled capability.

---

Source: Grafy Research Lab — AutoBuddy: Evaluating Proactive, Always-On Agents on Task Discovery and Execution in a Simulated Workplace