Advancing the Search Frontier with AI Agents

Task modeling: Build richer task models that more fully represent tasks and task contexts. This includes how we infer tasks and intent—for example, from the textual content of the search process, from user-system interactions, and from other situational and contextual information such as location, time, and application usage—and how we represent those tasks internally—for example, as a hierarchy (Figure 1) or a more abstract representation (semantic vectors, graph embeddings, Markov models, and so on). We also need to be able to estimate key task characteristics such as task complexity, which, in one use, can help search systems route requests to the most appropriate interaction modality. In addition, we need to find ways for agents to collect more, and more-accurate, user and world knowledge, both in general and specifically related to the task at hand. A better understanding of the short- and long-term task context will help agents more accurately model the tasks themselves.

Alignment: Develop methods to continuously align agents to tasks, goals, and values via feedback. Here, feedback includes conversation content, such as searchers expressing positive sentiments like gratitude to the agent, or explicit feedback on agent answers via likes and dislikes. The performance of agents that are lacking alignment will remain fixed over time. Agents need application-aligned feedback loops to better understand searcher goals and tasks and must use that feedback to continuously improve answer accuracy and relevance. Beyond research on fine-tuning foundation models from human feedback (for example, likes and dislikes),⁴⁵ we can also build on learnings from research on implicit feedback in IR, including work on improving ranking algorithms via SERP clicks¹⁶ and developing specialized interfaces to capture user feedback.³⁸

Augmentation: Enhance agents with relevant external knowledge and enhanced tools and capabilities. As mentioned earlier, RAG is a common form of knowledge injection for foundation models. Relevance models are tuned to maximize user benefit, not for agent consumption. We need to evaluate whether this difference is meaningful practically, and if so, develop new ranking criteria that consider the intended consumer of the search results (human or machine). Despite their incredible capabilities, foundation models still have shortcomings that manifest in the agents that use them. We need to understand these shortcomings through principled evaluation and find ways to leverage external skills or plugins to address them. Agents must find and recommend skills per task demands,³⁹ for example, invoking Wolfram for computational assistance. We can also integrate tool use directly into tool-augmented models, such as Toolformer,²⁵ that can teach themselves to use tools. Models of task context may also be incomplete, so we should invest in ways to better ground agent responses via context using, for example, richer sensing, context filtering, and dynamic prompting.

Grounding: Apply use-case-specific information to reduce hallucinations, build trust, and support content creators. It is in the interests of agents, searchers, and content creators (and providers and advertisers) to consider the source of the data used in generating answers. Provenance is critical. Agents should provide links back to relevant sources, preferably with specific details or URLs, to help establish and maintain user trust, provide attribution for content creators, and drive engagement for content providers and advertisers. To build trust and support learning, it is also important for agents to practice faithful reasoning¹⁰ and provide interpretable reasoning traces (for example, explanations with chain-of-thought descriptions) along with their answers. We should also consider how we can integrate agents within existing experiences (for example, in other applications) to ground answers in their context of use.

Personalization: Develop personal agents that can understand searchers and their tasks while using personal data, privately and securely. Searchers bring their personal tasks to search systems, as they will with agents. Here are some example personal prompts that describe the types of personal tasks that searchers might expect an agent to handle:

These tasks, spanning creation, summarization, and recommendation, quickly illustrate the wide range of expectations that people may have from their personal agents. As part of developing such personalized AI support, we need to do two key things. The first is to study foundation model capabilities, including their ability to identify task-relevant information in personal data and activity histories, and model user knowledge in the current task and topic. The second is to develop core technologies, including the following:

Adaptation. Two main forms of adaptation that we consider here are model specialization and so-called adaptive computation:

These adaptation methods will yield more-effective and more-efficient AI capabilities that agents can use to help searchers across a range of settings, including in offline settings (for example, on-device only).

Multi-agent: Utilize multiple specialized agents working together and with humans to help complete a search task. Multiple agents have been shown to help encourage divergent thinking, improve factuality and reasoning, and provide guardrails for AI systems. Multi-agent systems such as AutoGen⁴¹ orchestrate communication between agents to help users complete tasks more effectively. These systems could be used in search settings to, for example, retrieve relevant resources from diverse sources, improve answer correctness by critiquing and refining AI-generated output, improve human decision making by presenting alternative solutions, and even automate the completion of some tasks or sub-tasks, with humans in the loop throughout.

Multimodality: Develop experiences bridging (at least) the search and agent (chat) modalities, offering explanations and suggestions. Given how entrenched and popular traditional search is, it is likely that some form of query-result interaction will remain a core part of how we find information online. Future, agent-enhanced experiences may reflect a more-seamless combination of the two interaction modalities in a unified experience. Both Google and Bing are taking a step in that direction by unifying search results and agent answers in a single interface. Explanations on which modality and style (for example, creative, balanced, or precise) perform best and when will help searchers make decisions about which modalities and settings to use. Modality recommendation based on task is also worth exploring: Simple tasks may need only traditional search, whereas complex tasks may need agents. Contextualization and personalization will also play an important part in deciding how much information is needed from the searcher (incurring interaction cost but yielding greater control) and how much can be reliably inferred from signals already available to the system. Related to this are opportunities around conversation-style suggestion given the current task. For example, a simple fact-finding task needs short, precise replies (when generative-AI-powered agents can often be verbose), while generating new content needs creativity (when agent responses can often be unoriginal or bland). Search providers could also consider offering a single point of entry and an automatic routing mechanism to direct requests to the correct modality given inferences about the underlying task (see “Task modeling” section above) and the appropriateness of each of the modalities for that task. Beyond search and chat, other modalities to help support complex search tasks may include third-party tools and applications, bespoke user interfaces (for example, tailored dynamically by the agent to the task at hand), interactive visualizations, and proactive recommendations.

Human learning: Develop agents that can detect learning-related search tasks and support relevant learning activities. As mentioned earlier, agents can remove or change human learning opportunities by their automated generation and provision of answers. Learning is a core outcome of information seeking.^11,21,35 We need to develop agents that can detect learning and sensemaking tasks, and support relevant learning activities via agent experiences that, for example, provide detailed explanations and reasoning, offer links to learning resources such as instructional videos, enable deep engagement with task content via relevant sources, and support specifying and attaining learning objectives. A good example of all of this is Microsoft's recently announced partnership with the Khan Academy.^g

Human control: Better understand control and develop agents with control while growing automation. Control is an essential aspect of searcher interaction with agents. Agents should consult humans to resolve or codify value tensions. Agents should be in collaboration mode by default and must only take control with the permission of stakeholders. Experiences that provide searchers with more agency, such as adjusting the specificity or diversity in agent answers, are critical, leading to less generality and less repetition. As mentioned in the “Grounding” section above, citations in answers are important. Humans need to be able to verify citation correctness in a lightweight way, ideally without leaving the user experience; Gemini now offers an ability to manually dive deeper and verify answers. We also need a set of user studies to understand the implications of providing less control over some aspects, such as answer generation; more control over other aspects, such as macrotask specification; and control over new aspects, such as conversation style and tone (as with Copilot).

Completion: Agents should help searchers complete tasks while keeping searchers in control. We need to both expand the search frontier by adding or discovering more capabilities of foundation models that can be surfaced through agents and deepen task capabilities so that agents can help searchers better complete more tasks. We are moving from a world equipped with only search engines, to one also equipped with AI-powered answer engines, with agents that provide relevant information, synthesized from several sources, and action engines that can perform actions to complete tasks, or help find agents to do so. We can view skills and plugins as actuators of the digital world and should help foundation models fully utilize them. We need to start simple (for example, reservations), learn and iterate, and increase task complexity as model capabilities improve with time.

The standard mode of engagement with AI agents is reactive; users send requests and the agents respond. Agents should ideally have a dynamic interaction model that tailors the interface to the task and the context. With this model, agents can take initiative, with permission, and provide updates for standing tasks, or they can offer proactive suggestions or take actions directly when agent uncertainty is low. Agents can also help support task planning (decomposition, prioritization, and so on) for complex tasks such as travel or events. AI can already help complete repetitive tasks (for example, action transformers trained on digital tools^h) and create and apply “tasklets” (user interface scripts) learned from websites.¹⁹

Given the centrality of information interaction in search task completion, it is important to focus sufficient attention on interaction models and experiences in AI agents. In doing so, we must also carefully consider the implications of critical decisions on issues that affect AI in general, such as control and automation.

Understanding: Deeply understand agent capabilities and agent impact on searchers and their tasks. We have only scratched the surface in understanding AI agents and their impact. Gaining a deeper understanding could take a few forms. The first is user understanding, covering the mental models of agents and the effects of bias (for example, functional fixedness) on how agents are adopted and used in search settings. It also covers changes in search behavior and information seeking strategies, including measuring changes in effects across modalities, for example, search versus agents and search plus agents. There are also opportunities in using foundation models to understand search interactions via user studies and generate intent taxonomies and classify intents from log data. The second form is task understanding, covering the intents and tasks for which agents are used and most effective. Finally, there is agent understanding, covering the capabilities and limitations of agents. This form of inquiry is similar to that found in the “Sparks of AGI” paper on GPT-4,⁷ which examined foundation model capabilities in depth.

Evaluation: Identify and develop metrics for agent evaluation, while considering important factors, and find applications of agent components for IR evaluation. There are many options for AI agent metrics, including feedback, engagement, precision-recall, generation quality, and answer accuracy. Given the task focus, metrics should likely target the task holistically (for example, success, effort, satisfaction). In evaluating agents in search settings, it is also important to consider the following:

There are also opportunities to consider applications of agent components for IR evaluation. Foundation models can predict searcher preferences³³ and assist with relevance judgments,¹² including generating explanations for judges. Also, foundation models can create powerful searcher simulations that can mimic human behavior and values, expanding on early work on searcher simulations in IR.

Measuring agent performance is essential in understanding their utility and improving their performance over time. Agents do not function in a vacuum, and we must consider the broader implications of their deployment for complex tasks in search settings.

Responsibility: Understand factors affecting reliability, safety, fairness, and inclusion in agent usage in search. The broad reach of search engines means that AI agents have a critical obligation to act responsibly. Research is needed on ways to improve answer accuracy via better grounding in more-reliable data sources; developing guardrails; understanding biases in foundation models, prompts, and the data used for grounding; and understanding how well agents work in different contexts, with different tasks, and with different people and cohorts. Red teaming, user testing, and feedback loops are all needed to determine emerging risks in agents and the foundation models that underlie them. This also builds on existing work on responsible AI, responsible IR, and FACTS-IR, which has studied biases and harms and ways to mitigate them.²⁴

Economics: Understand and expand the economic impact of agents in search. This includes exploring new business models that agents will create beyond information finding. Advancing the search frontier from information finding deeper into task completion (for example, into creation and analysis) creates new business opportunities. It also unlocks new opportunities for advertising, including advertisements that are shown inline with dialogue or answers and are contextually relevant to the current conversation. There is also a need to more deeply understand the impact of agents on content creation and search engine optimization. Content attribution is vital in such scenarios to ensure that content creators (and advertisers and publishers) can still generate returns. We should avoid the so-called paradox of reuse,³⁶ where fewer visits to online content leads to less content being created, which in turn leads to worse models over time. Another important aspect of economics is the cost-benefit trade-off, related to work on adaptation (see “Adaptation” section above). Large model inference is expensive and unnecessary for many applications. This cost will reduce with optimization, for which model specialization and adaptive computation can help, as does the emergence of high-performing SLMs of a few billion parameters, such as Phi, trained on highly curated data.¹⁴

Ubiquity: Integrate agents to model and support complex search tasks. AI agents must coexist with the other parts of the application ecosystem. Search agents can be integrated into applications such as Web browsers, offering in-browser chat, editing assistance, and summarization; and productivity applications, offering support in creating documents, emails, presentations, and so on. These agents can capitalize on the application context to do a better job of answering searcher requests. Agents can also span applications through integration with the operating system. This enables richer task modeling and complex task support, since such tasks often involve multiple applications. Critically, we must do this privately and securely to mitigate risks for searchers and earn their trust.