Joint survey by UIUC, Google, and others: Toward superintelligent agents, what kind of proactivity does AI need?

In 1995, Michael Wooldridge and Nicholas Jennings, in "Intelligent Agents: Theory and Practice," listed autonomy, social ability, reactivity, and proactiveness as the four basic characteristics of agents.

Proactiveness means that an agent can actively work toward goals.

This requirement entered the basic discussion of agent research thirty years ago.

Today, large-model agents can already understand instructions, call tools, and independently complete tasks that involve multiple steps. Many systems still require a person to identify problems, propose tasks, and then hand them to AI for processing. People's expectations for intelligent assistants also include another capability: before new instructions arrive, to discover work worth doing and proactively set about handling it.

Figure 1 | Research lineage supporting proactive behavior. Continuous perception, prediction, goal maintenance, and intervention selection all have research foundations that predate large models.

This capability directly affects whether long-horizon tasks can continue to advance. When a piece of work lasts for days or weeks, the initial instructions can hardly specify every change that occurs later. Information may be missing, the original plan may be blocked, and arrangements already completed may need to be re-examined. Agents need to proactively clarify questions, seek alternatives, and adjust plans in order to reduce the burden on people of repeatedly checking progress and supplementing instructions.

Imagine a researcher delegating conference travel and subsequent changes to an assistant, with flights and hotels already booked. One week before departure, the conference is extended by a day, so the return flight is now earlier than the conference end time, and the hotel's free cancellation deadline falls on that same evening. Continuing with this trip requires the assistant to discover the conflicts among these arrangements, verify the conditions for adjustment, prepare options, and promptly raise the matters that require the researcher's decision. The instructions given at the time of booking can hardly specify in advance how this chain of changes should be handled.

For future superintelligent agents, proactiveness should be valued alongside intelligence and alignment. Intelligence concerns the ability to understand, reason, and execute; alignment concerns whether behavior conforms to human intent and values. Proactiveness also requires an agent to judge what is worth doing and to bear responsibility for discovering and initiating work. An assistant capable of long-term collaboration needs capabilities in all these respects.

Frontier labs have already begun exploring such assistants. Meta positions Muse as a personal agent that can proactively offer help around people's goals and continuously advance work; OpenAI's dots supports continuous follow-up on tasks across conversations and can also proactively research from information it is permitted to read, discovering new situations related to existing work.

Academia is also studying concrete implementation approaches. The Proactive Agent from Tsinghua University and other institutions studies how to predict tasks based on context and proactively offer help when there are no explicit instructions; ProAgent from the Chinese University of Hong Kong and other institutions uses wearable devices to obtain environmental information, infer needs in daily activities, and call tools to provide assistance.

How to discover needs, select work worth undertaking, and then keep it moving forward is a problem these explorations all face. In "Proactive AI for Superintelligent Agents," researchers from UIUC, Google, Stanford, UC Berkeley, and CMU, focusing on the capability of proactively initiating work, propose the 6A analytical framework, organizing the relevant mechanisms, evaluation methods, and 13 application domains. Readers can use this framework to understand existing systems and to consider what other capabilities a proactive agent still needs.

What judgments does an agent need to make on its own

When querying flights and completing a booking as requested, the work to be done has already been pointed out by a person. After the conference is extended, the assistant needs to discover on its own that the return trip and accommodation are affected, judge which arrangements are worth re-examining, and propose follow-up work.

The goal of attending the conference and the budget still apply; which specific matters need to be handled must be judged according to the changed situation. This ability to discover and initiate work on one's own is the focus when the survey discusses proactiveness.

The survey groups these judgments into two questions: what is worth handling, and whether, when, and how to intervene. The return-trip conflict affects the goal of attending the conference, so there is a clear reason to check the return arrangements.

After discovering the conflict, the assistant also needs to compare the benefits and costs of adjustment, consider the cancellation deadline, and decide whether it is currently appropriate to prepare options, request confirmation, or execute operations that have already been approved.

Figure 2 | Reactive and proactive behavior. A proactive agent judges on its own what is worth handling and whether, when, and how to intervene; explicit requests and preset rules can still play a role in the same system.

An instruction can specify the conference dates, the budget cap, and the accounts permitted for use. How to trade off transfer risk against ticket price, how to adjust when accommodation cannot be extended, and which existing arrangements are worth keeping are often left to be judged at execution time. The paper uses "residual decision space" to describe these choices that are not specified in advance. Proactiveness can be used both to complete the current task and, under ongoing responsibilities, to discover new concrete work; human goals and constraints provide the basis for these choices.

Figure 3 | Some requirements in a task are already explicit, while others are left to be judged at execution time. An agent can handle these residual decisions within the scope delegated to it, and the goals and constraints continue to apply.

The ability to initiate work and the authority to execute it can be arranged separately. The assistant proactively identifies itinerary conflicts and prepares rebooking options, while payment can be left for the researcher to confirm; once the researcher explicitly requests rebooking, the assistant can also independently complete operations that have already been authorized. Proactiveness concerns how work is discovered and proposed, while the degree of autonomy concerns how far the system can execute independently. Intelligence supports its understanding and problem-solving, while alignment concerns whether behavior conforms to human intent and values.

Figure 4 | Proactiveness, intelligence level, and alignment jointly influence agent performance. The degree of autonomy describes the scope of authorized independent execution, and authorization needs to reference reliability and task risk.

The same assistant can adopt multiple working modes. Answering flight information after receiving a query request, sending a boarding reminder at the scheduled time, and proactively checking the itinerary based on a new meeting arrangement can all appear within one system. As the responsibilities the system undertakes change, humans can also shift from item-by-item operation to jointly revising plans, and then to reviewing key decisions. Responsive behavior, proactive judgment, and the mode of human participation need to be understood together in the context of specific tasks.

Figure 5 | Collaboration modes can change continuously. Tasks, evidence, time scales, and permissions jointly influence the agent's degree of proactiveness and the mode of human participation.

From a rescheduling notice

to a proactive assistance

In the itinerary example, self-discovering work depends on several mutually coordinated capabilities. Notifications and orders provide the current situation, anticipation helps identify downstream impacts, agenda selection determines which matters to follow up on, arbitration judgment determines when to advance, and execution and feedback allow the work to continue. The survey groups these capabilities into 6A: Awareness, Anticipation, Agenda, Arbitration, Action, and Adaptation.

Large language models and multimodal models can process text, images, and other contextual information. The memory, tools, skills, and execution management provided by the Agent harness extend the agent's working capabilities, enabling it to connect to external applications, save progress, and continue tasks across sessions. The 6A framework helps explain how these capabilities are used to discover and advance work, and how different stages influence one another.

Figure 6 | The 6A framework. The family scenario in the figure illustrates how the capabilities influence one another, and how execution results and human feedback can revise prior understanding and arrangements.

Awareness

Checking new notifications against existing arrangements

The new meeting end time, the originally scheduled return time, and the hotel cancellation deadline are scattered across notifications, flight tickets, and hotel orders. The assistant associates this information with the current goal of attending the conference and updates its understanding of the itinerary. At this point, the meeting extension has been confirmed, while the flight and hotel are still booked for the original dates, giving subsequent judgments a clear starting point.

Cross-session work also depends on the preservation and updating of information. Memory retains the conference attendance goal, budget, and completed arrangements, and new notifications change the dates within them. When rebooking terms are missing, the assistant can proactively query them; when resuming work again, it can carry over records that remain valid. Awareness covers this kind of state tracking, information acquisition, and memory management, and also includes how the system acquires relevant changes when there is no new question.

Figure 7 | When work spans multiple sessions, state and memory need to remain consistent. Continuous observation also involves privacy, computation, and energy costs, and the scope and frequency of information acquisition need to be commensurate with the task.

Anticipation

Judging what consequences a change will bring

With the updated itinerary, the assistant can infer that the original return flight will conflict with the meeting on the last day. The postponed return will also create accommodation needs, which may affect the budget and subsequent schedule. The approaching hotel cancellation deadline also makes waiting incur potential costs. Anticipation turns a date change into work needs that have not yet been raised by a human.

These judgments involve both human goals and external conditions. The conference attendance goal determines which time conflicts need to be handled, and the budget and personal arrangements influence the trade-offs among options; flight schedules, remaining seats, rebooking rules, and hotel terms determine whether an option is feasible. The user model and world model in the paper respectively help the system understand this information and predict subsequent needs and environmental changes.

The assistant can also compare the consequences of keeping the original arrangements, postponing the return, or choosing other options, and estimate what benefit early intervention can bring. Such comparison requires distinguishing verified facts from assumptions in predictions. Speculation about remaining seats or costs can lead to a query, and the query results are then used to revise the plan.

Figure 8 | Anticipation considers both human needs and environmental changes. Comparing action and waiting helps judge what benefit early intervention can bring.

Agenda

Selecting and continuously following up on work worth doing

After discovering that the return, accommodation, and budget are affected, the assistant selects which matters enter the to-do list and organizes them into work that can be advanced.

The to-do list can include querying rebooking and extended-stay terms, preparing a cost comparison, and raising matters that require the researcher's decision. The free cancellation deadline affects the order of handling, and flight choice also affects accommodation arrangements.

Agenda concerns goal selection, prioritization, and continuous maintenance. Storing the basis, progress, and dependencies of each piece of work in the to-do list enables the assistant to continue following up after waiting for confirmation or after a cross-session interruption. If the extended-stay terms have been clarified and the return flight has not yet been selected, the next resumption can directly handle the unfinished part. When new important tasks arrive, existing work can also be reprioritized while retaining the information needed to resume later.

Figure 9 | Agenda maintains existing commitments and adjusts priorities as new requests and environmental changes arise. Predicted needs must go through selection before they can become goals that are continuously followed up on.

Arbitration

Judging which step to advance to at this moment

The return trip and accommodation have already entered the to-do list, and the assistant still needs to decide in what manner to proceed now. Organizing options, reminding about deadlines, requesting confirmation, and directly modifying the booking each produce different effects. This step combines evidence, benefits, timing, and risk to decide whether to intervene, as well as the timing and extent of intervention.

In this trip, the meeting dates have been confirmed and the hotel deadline is also known, so there is a clear reason to prepare a plan in a timely manner. The specific costs and extended-stay conditions still need to be checked, and payment and cancellation operations are still constrained by permissions. The assistant can first complete information queries and plan organization, promptly raise the matters that require the researcher's decision, and handle the booking only after confirmation.

The judgment here also includes uncertainty and the degree to which actions can be withdrawn. An adjustment may resolve a scheduling conflict while incurring high costs; after canceling a booking, the original price or room may also be impossible to restore. Arbitration incorporates these consequences into the choice of intervention, helping the system turn "a problem has been discovered" into a justified current action.

Figure 10 | Intervention judgment considers evidence, value, timing, and risk together to decide which step is appropriate to advance to at present.

Action

Turning decisions into understandable assistance

Once the intervention approach is chosen, the assistant can organize the original booking, new dates, alternative flights, cost differences, and deadlines into a plan for the researcher to compare and modify. At this step, how information is presented affects the time required for review. The researcher can clearly see which arrangements are being modified, where the costs come from, and which operations have not yet taken effect.

After the researcher selects a plan and confirms the operations, the assistant can complete the rebooking or accommodation adjustment within the authorized scope, and then verify the results.

When operations involve multiple applications, some steps may succeed while others fail, so existing bookings and pending matters need to be recorded separately. The paper includes suggestions, inquiries, drafts, previews, and direct execution all within Action, and also discusses how editing, undoing, human takeover, and failure recovery support collaboration.

Figure 11 | The intervention approach determines how humans participate. Previews, confirmations, modifications, undos, and failure recovery make specific operations understandable, reviewable, and correctable.

Adaptation

Adjusting subsequent work based on results and feedback

After the trip is adjusted, tool results can indicate whether the rebooking succeeded, while the researcher's modifications provide feedback on the plan. When an operation fails, the assistant can re-query or adjust its execution approach based on the error. If the researcher modifies the return option, the assistant can also update the current arrangements accordingly. Feedback thus influences subsequent state understanding, goal maintenance, and action.

Cross-task learning also requires interpreting the scope of applicability of feedback. If the researcher replaces the suggested late flight with an early one, the reason may be another commitment that day, or it may be a long-term preference. A single choice can correct the current trip, but its significance for future business trips requires more evidence. Failing to respond to a reminder may likewise be related to being busy, the timing of the prompt, or its content.

Adaptation focuses on corrections in the current task, and also on personalization and adjustment of proactivity in long-term collaboration. What the system needs to learn includes needs judgment, reminder timing, and intervention approach. Whether execution succeeds, whether suggestions are accepted, and whether proactive assistance aligns with human goals provide feedback with different meanings.

Figure 12 | Feedback has delays and noise. The system needs to interpret the meaning of signals before deciding how to correct current work, update preferences, or adjust subsequent collaboration approaches.

When new situations arise during execution, the assistant can re-acquire information, adjust goals, or modify the plan. Feedback also affects subsequent predictions and intervention judgments, so the six capabilities repeatedly coordinate with one another in ongoing work.

13 application domains: where proactivity is being used

Itinerary arrangement belongs to personal assistance, and the survey also examines coding, graphical interface operation, information acquisition, embodied robots, security, automated scientific research, smart homes, smart cities, healthcare, law, education, entertainment and media, and agriculture, 13 domains in total. Comparing these applications reveals how the need to proactively discover work varies with tasks, environments, and human responsibilities.

Figure 13 | The paper examines 13 application domains. The home scenario on the right depicts a future direction: humans set goals, and agents continuously discover needs, arrange work, and accept human review and feedback.

In the coding domain, the focus is on how to discover unassigned problems from code, tests, and usage feedback, propose justified fixes, and follow up on maintenance work.

Some systems can already prepare fixes and test results, while continuously maintaining an entire project requires selecting which issues are worth addressing, setting priorities, and preserving unfinished work. Automated scientific research systems propose hypotheses around directions given by researchers, select experiments, and then arrange subsequent research based on results; the judgments they face also include whether a problem and experiment are worth investing in.

Proactive assistance in home scenarios is often related to daily routines. Existing systems can remind about forgotten device operations or predict heating needs based on usage patterns. Continuously handling household affairs also involves the differing preferences of multiple family members, as well as whether a reminder or operation is appropriate at the time. Urban systems use predictions to assist decision-making in public affairs such as traffic, while specific actions are usually still decided by the relevant personnel and institutions.

The paper separately records the degree of proactivity and execution authority of these systems. Some scientific research systems can independently execute longer processes, while high-level research questions are still given by humans; home systems can issue reminders on their own, and the scope of device operations is specified separately. Among the systems surveyed, long-term maintenance of established responsibilities still has room for development, and the ability to revise higher-level goals on one's own is not yet mature.

Cross-domain comparisons reveal three trends: foundation models and domain-specific methods jointly take on the work, some agents are beginning to take charge of processes that include multiple stages, and the reliability of feedback and the cost of verification affect continuous improvement.

Figure 14 | Three cross-domain trends. Sustained assumption of work requires general reasoning, domain capabilities, and verifiable results; the specific form depends on application conditions.

A single act of proactive assistance

What exactly was saved

If the evaluation directly states "the meeting has been rescheduled, please modify the itinerary," the problem to be handled has already been pointed out by the tester. When examining proactive capability, it is also necessary to hand the official notice, existing orders, and current responsibilities to the agent and let it determine whether there is work worth initiating. What needs were discovered, what was missed, and what unnecessary changes were proposed should all be recorded.

The survey organizes evaluation at three levels: task, mechanism, and user. In the business trip example, the task level checks whether the return trip is scheduled after the meeting ends, whether the accommodation covers the meeting period, whether the budget is exceeded, and whether duplicate orders or cancellation losses occur. The time, computation, and human input required to complete this work should also be counted as costs. Only in this way can it be determined how much practical benefit early handling brought.

The mechanism level checks whether the assistant identified the official rescheduling notice, found the relevant orders, correctly predicted the schedule conflict, and arranged the matters requiring confirmation before the deadline.

The timing of suggestions, the evidence required, and the permissions are likewise to be checked. Continuing verification when the date has not yet been determined, and stopping follow-up when the need has disappeared, can both be counted as reasonable choices in the evaluation. Errors should be traceable to the information, prediction, goal selection, intervention judgment, or execution stage.

The user level records how much time researchers spend checking plans, supplementing information, and correcting errors, whether reminders interrupt work, whether order status is easy to understand, and whether modification, rejection, or stopping operations are convenient. In long-term use, it is also necessary to examine whether people can decide when to adopt and when to review based on the assistant's actual reliability. Both the work reduced by proactive assistance and the added review burden must be supported by evidence.

Figure 15 | Three-level evaluation connects work outcomes, the mechanisms that generate actions, and people's user experience. Human review and correction input is also a cost of proactive assistance.

Continuing the trip a few days later, going on another business trip to a different destination, or revising a long-term budget can all test problems that are difficult to observe in short-term tests. Whether the assistant carries over old preferences, whether it can resume interrupted work, and whether it re-verifies after information becomes outdated require tests spanning weeks and months. The paper also proposes detailed mechanism-level checks to help determine where long-term performance problems specifically arise.

How proactivity supports long-horizon tasks and self-evolution

In long-horizon tasks such as scientific research, the next step often depends on new discoveries. Hidden constraints may invalidate the original plan, and a single failure may also reveal new research questions; even human experts may not be able to provide a complete path in advance.

The paper argues that proactivity is a necessary capability for agents to cope with such tasks: to notice problems in the current path on their own, propose new goals worth exploring, and proactively clarify when needed, so that work can continue to advance without next-step instructions.

Proactivity also provides direction for self-evolution. Agents can discover capability bottlenecks from unmet needs, anomalies, and their own mistakes, and proactively decide what to learn next and what to explore.

What is learned can in turn be used to improve understanding of situations, predictions of the future, and goal selection and intervention judgment. The paper proposes that research should examine how these two capabilities reinforce each other, so that self-evolution improves the entire 6A process.

Self-generated goals and capability updates both need to be subject to oversight. Why an agent chose a certain goal and what changes occurred in its behavior after learning should be inspectable and correctable, and rollback should be supported as much as possible. Human authorization and external safety checks also need to constrain what agents pursue and how they improve themselves.

Toward superintelligent agents, intelligence, alignment, and proactivity need to develop together.

Agents with these capabilities are expected to become partners capable of long-term collaboration: continuously discovering work worth doing, proposing exploration directions when encountering unknown problems, and improving themselves in practice.

Such collaboration is expected to reduce the burden on people to repeatedly monitor the environment, supplement instructions, and adjust plans, allowing people to devote more energy to deciding which problems are worth studying and which goals are worth pursuing, and to explore questions without ready-made answers together with agents.

References:

This article comes from the WeChat public account "Xinzhiyuan"; author: Xinzhiyuan; editor: LRST; published by 36Kr with authorization.