Delegating to Machines: What Early AI Agent Users Reveal

by

Author: Denis Avetisyan

New data from the Perplexity AI agent, Comet, provides a first look at how people are actually using this emerging technology to automate tasks and augment their abilities.

Agentic queries, as demonstrated by Perplexity, showcase a capacity for complex information seeking that extends beyond simple keyword matching, revealing an ability to formulate and refine requests to achieve nuanced results.
Agentic queries, as demonstrated by Perplexity, showcase a capacity for complex information seeking that extends beyond simple keyword matching, revealing an ability to formulate and refine requests to achieve nuanced results.

This study analyzes behavioral data to characterize early adopters, identify dominant use cases-primarily productivity and learning-and understand patterns of task delegation to AI agents.

Despite the rapid proliferation of artificial intelligence, understanding who is adopting AI agents and how they are being utilized remains largely unexplored. This paper, ‘The Adoption and Usage of AI Agents: Early Evidence from Perplexity,’ presents a large-scale analysis of user interactions with Comet, an AI-powered browser and its integrated agent, revealing that early adopters are concentrated among digitally-savvy professionals in knowledge-intensive fields, primarily leveraging the agent for productivity and learning tasks. These users exhibit consistent engagement with specific use cases, though a shift towards more cognitively demanding applications emerges over time. What implications do these early patterns of adoption and usage hold for the future development and societal integration of increasingly capable AI agents?

The Illusion of Intelligence: From Reactive Systems to Autonomous Agents

Conventional artificial intelligence systems often falter when confronted with tasks demanding a sequence of independent decisions and actions. Historically, AI has excelled at narrowly defined problems – image recognition, data analysis – but struggles with the adaptability required for complex, real-world scenarios. These systems typically require explicit, step-by-step programming for each possible contingency, a limitation that renders them brittle and inefficient when faced with unforeseen circumstances. The inability to autonomously plan, execute, and refine strategies across multiple stages necessitates constant human intervention, hindering their potential for true automation and scalability. This reliance on pre-defined pathways sharply contrasts with human problem-solving, where iterative thought and flexible action are commonplace, and highlights the need for a new approach to artificial intelligence.

Agentic AI systems signify a fundamental departure from traditional artificial intelligence, moving beyond mere response to requests towards genuine autonomous action. These systems aren’t simply programmed to react; they are designed to independently pursue objectives established by the user, effectively functioning as digital agents. This capability stems from an ability to decompose complex goals into manageable steps, plan a course of action, and iteratively execute that plan, learning and adapting along the way. Unlike conventional AI which requires explicit, step-by-step instructions, agentic systems exhibit a degree of proactivity, proactively seeking information and utilizing tools to achieve the desired outcome, mirroring a level of cognitive flexibility previously unseen in artificial intelligence and promising a future where technology anticipates and fulfills needs with minimal human intervention.

The emergence of agentic AI, leveraging the capabilities of large language model chatbots, signals a fundamental shift in the human-technology relationship. Historically, interaction demanded explicit instruction for each step of a task; now, these systems are designed to autonomously pursue defined objectives, effectively acting as digital collaborators. This transition moves beyond simple question-and-answer exchanges towards a paradigm where users articulate goals, and the AI independently formulates plans, executes them through various tools and APIs, and iteratively refines its approach-all without constant human intervention. The implications are vast, potentially automating complex workflows, personalizing experiences at an unprecedented level, and fundamentally altering how individuals engage with digital services, moving from directing tools to collaborating with intelligent agents.

The advancement of agentic AI hinges significantly on frameworks such as ReAct, which facilitate a dynamic cycle of thought and action. Unlike traditional AI models that execute pre-defined sequences, ReAct enables systems to iteratively reason about a goal, formulate a plan, take an action, and then observe the outcome before revising its approach. This process, mimicking human problem-solving, allows the AI to navigate complex tasks with greater flexibility and resilience. By interweaving reasoning traces with action steps, the framework not only improves performance but also offers a degree of transparency into the AI’s decision-making process. Consequently, ReAct and similar frameworks are proving instrumental in building AI agents capable of autonomous operation and adaptation in real-world scenarios, marking a substantial leap beyond static, rule-based systems.

The agentic taxonomy is structured hierarchically, organizing agents based on levels of complexity and capability.
The agentic taxonomy is structured hierarchically, organizing agents based on levels of complexity and capability.

Comet: A Watcher in the Machine Room

Comet, developed by Perplexity, functions as a dedicated platform for the observation and analysis of agentic AI systems. This platform provides an environment where researchers can deploy and monitor autonomous agents as they interact with various tools and data sources. Its core function is to enable systematic study of agent behavior, facilitating the collection of data on task completion, decision-making processes, and overall system performance. This capability is particularly valuable for understanding the emergent properties of complex agentic systems and evaluating their reliability and safety. The platform’s design emphasizes observability, allowing for detailed tracking of agent actions and internal states, which is critical for identifying potential issues and improving agent design.

Comet utilizes the Perplexity platform’s existing infrastructure to provide a functional environment for agentic AI systems to interact with external tools and APIs. This integration allows researchers to not only deploy agents but also to monitor and analyze their actions in real-time. The Perplexity platform handles the underlying complexities of API access, data retrieval, and execution, enabling a focused study of agent behavior. Specifically, Comet leverages Perplexity’s search capabilities and information access to provide agents with the necessary context for decision-making and task completion, while simultaneously logging all interactions for detailed analysis.

The Model Context Protocol (MCP) is a foundational element of the Comet platform, enabling agentic AI systems to interact with and utilize external applications. This protocol defines a standardized interface for exchanging information between the AI agent and external tools, allowing agents to move beyond simple text-based interactions and perform actions in the real world. Specifically, MCP manages the transmission of both input parameters to applications and the reception of resulting outputs, providing a structured data flow. This capability is crucial for extending the functionality of agentic AI beyond its core language model and facilitating complex, multi-step reasoning processes that require external data or action execution.

Comet facilitates the observation of agentic AI system behavior, providing researchers with data to assess their capabilities. Usage statistics demonstrate significant post-launch engagement; 60% of users adopting agent functionality and 50% of all agentic queries originated after the platform’s general availability. This indicates a rapid rate of adoption and active utilization of agentic features within the Perplexity platform, suggesting substantial interest in and exploration of agentic AI by the user base.

The plot illustrates the cumulative growth of both agent adoption and agentic queries over time.
The plot illustrates the cumulative growth of both agent adoption and agentic queries over time.

Where Are These Things Actually Being Used?

Agentic AI applications are currently distributed across several key domains, notably Learning, Media, and Productivity. Within the Learning sector, these agents facilitate personalized educational experiences and automated content creation. In Media, agentic AI powers content summarization, automated journalism, and enhanced content recommendation systems. The Productivity domain utilizes agentic AI for task management, automated scheduling, and streamlined communication workflows. These applications demonstrate the versatility of agentic AI and its potential to augment human capabilities across a broad spectrum of professional and personal activities.

The Agent Usage Ratio (AUR) serves as a key performance indicator for understanding how frequently agentic AI tools are utilized by adopted users. Calculated as the average number of agent interactions per user within a defined period, the AUR differentiates between simple adoption and active engagement. A higher AUR indicates that users are not merely accessing the technology, but are consistently integrating it into their workflows. Current data reveals that the Digital Technology and Entrepreneurship sectors demonstrate an AUR of 1.1-1.2, signifying a robust level of usage intensity following initial adoption, and providing valuable data points for assessing the return on investment and identifying best practices in agentic AI deployment.

The Agent Adoption Ratio (AAR) serves as a key performance indicator for gauging the extent of agentic AI integration within user workflows. Currently, the Hospitality sector exhibits the highest AAR, registering at 1.36. This value indicates that, on average, users within the Hospitality sector are demonstrating adoption levels exceeding one agent per user, suggesting a substantial integration of agentic AI tools into their standard operating procedures. Comparative analysis of AAR across different sectors provides actionable insights into the varying degrees of agentic AI implementation and potential areas for focused development and deployment.

Analysis of Agent Usage Ratio (AUR) and Agent Adoption Ratio (AAR) provides quantifiable data for evaluating the success of agentic AI applications and informing future development efforts. Current data indicates that the Digital Technology and Entrepreneurship sectors demonstrate a higher AUR, ranging from 1.1 to 1.2. This suggests that, within these sectors, users are not only adopting agentic AI tools but are also utilizing them with greater intensity – meaning they are leveraging the capabilities of these agents more frequently and extensively once integrated into their workflows – compared to other sectors.

Agentic queries demonstrate a distinct topic distribution compared to non-agentic ones.
Agentic queries demonstrate a distinct topic distribution compared to non-agentic ones.

The Future of Work: Or, How to Explain Your Job to the AI

The Digital Technology Career Cluster is experiencing a surge in novel roles directly attributable to the rise of agentic artificial intelligence. These aren’t merely existing positions rebranded; rather, agentic AI is catalyzing demand for specialists in prompt engineering, AI model customization, and the ethical oversight of autonomous systems. Furthermore, opportunities are burgeoning in areas focused on integrating these intelligent agents into existing workflows, requiring professionals adept at change management and human-AI collaboration. This expansion isn’t limited to highly technical roles; positions demanding creative problem-solving to define agentic AI’s application, and critical analysis to validate outputs, are also gaining prominence. The resulting landscape showcases a dynamic shift, promising a robust future for skilled professionals within the digital technology sector who can effectively harness the power of agentic AI.

The ONET database, a comprehensive resource detailing occupational characteristics and skill requirements, offers a valuable lens through which to examine the evolving impact of agentic AI across diverse professions. By categorizing work into detailed clusters – such as those focused on business, management, and administration, or science, technology, engineering, and mathematics – ONET allows for a granular assessment of how these intelligent agents are reshaping job roles. Analysis reveals that agentic AI isn’t uniformly disrupting all occupations; rather, it’s selectively augmenting tasks within specific clusters, particularly those involving information processing, data analysis, and routine cognitive work. This framework demonstrates that the future of work isn’t solely about job replacement, but rather a significant redefinition of roles, with humans and AI collaborating to achieve greater efficiency and innovation within established occupational structures. Understanding these shifts through the O*NET clusters is crucial for targeted skills development and workforce planning.

The integration of agentic AI signifies a fundamental shift beyond simple task automation, instead fostering a collaborative dynamic between humans and machines. Rather than replacing workers, these technologies are designed to augment human capabilities, handling repetitive or data-intensive aspects of jobs while freeing individuals to focus on strategic thinking, creative problem-solving, and complex interpersonal interactions. This isn’t merely about increased efficiency; it’s a reshaping of work itself, demanding a re-evaluation of skill sets and workflows. Consequently, professions are evolving to prioritize uniquely human strengths – critical analysis, emotional intelligence, and adaptability – as AI handles the more procedural elements. The resulting landscape isn’t one of job displacement, but of job transformation, requiring ongoing learning and a proactive approach to harnessing the power of agentic tools.

The evolving landscape of work, increasingly influenced by agentic AI, demands proactive skills development and workforce adaptation. Current data indicates a concentrated impact, with 96% of agentic query activity occurring within professional networking environments-suggesting these technologies are primarily being leveraged for career advancement and professional tasks. Notably, a significant 55% of these queries center around the top ten most common job functions, implying a focus on optimizing existing roles rather than entirely replacing them. This pattern highlights the necessity for individuals to cultivate skills in areas like prompt engineering, data analysis, and critical thinking to effectively collaborate with these intelligent agents, ensuring a seamless transition into a future where human capabilities are augmented, not diminished, by artificial intelligence.

The topic transition matrix reveals how agentic queries shift focus between successive interactions.
The topic transition matrix reveals how agentic queries shift focus between successive interactions.

The study of Perplexity’s Comet reveals predictable patterns. Early adopters, naturally, gravitate towards productivity and learning – leveraging the agent for task automation as a force multiplier. But this isn’t innovation; it’s simply applying existing need to a novel interface. The bug tracker will, inevitably, fill with requests for edge case handling and integration with legacy systems. As David Hilbert observed, “We must be able to answer the question: what are the ultimate foundations of mathematics?” This pursuit, mirrored in the rush to adopt AI agents, assumes a pristine theoretical space. Yet, production always introduces the grit. The elegant theory of agentic workflows will, inevitably, become tomorrow’s tech debt. It doesn’t deploy – it lets go.

What’s Next?

The observed enthusiasm for delegating tasks to agents, as demonstrated by Perplexity’s user base, feels less like a paradigm shift and more like the predictable expansion of surface area for failure. Every automation introduces a new class of error, exquisitely tailored to the specific anxieties of those attempting it. The current focus on productivity and learning, while pragmatic, obscures the inevitable creep toward more complex – and therefore more brittle – applications. One anticipates a future dominated not by elegant solutions, but by elaborate workarounds for systems that promised simplicity.

Future research must address the inherent opacity of these agentic systems. Behavioral data, while useful for identifying what users are doing, offers little insight into why they believe it works. The illusion of understanding is a powerful force, and it will likely sustain adoption long after the marginal benefits have diminished. A more critical line of inquiry should examine the cognitive biases at play – the tendency to anthropomorphize, to overtrust automated recommendations, and to conflate correlation with causation.

Ultimately, this field will be defined not by innovation, but by maintenance. Each layer of abstraction demands another layer of debugging. CI is the temple – one prays nothing breaks. The taxonomy of agents will inevitably fracture into a thousand competing implementations, each optimized for a niche use case and plagued by its own unique vulnerabilities. Documentation is a myth invented by managers, and the long tail of technical debt will stretch into the foreseeable future.

Original article: https://arxiv.org/pdf/2512.07828.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2025-12-10 03:05