Perplexity Comet's Declining Agent Capabilities: Why This AI Browser Is Becoming Timid

From "All-Powerful Assistant" to "Constant Deflection"

Recently, a Perplexity paid subscriber shared their experience on Reddit, sparking widespread discussion. The user mentioned they had been subscribed to Perplexity for two years and heavily used its AI browser product, Comet, to complete various tasks—filling out forms, completing compliance training, browsing Facebook Marketplace listings, and more. However, in recent weeks, these once-smoothly-functioning features suddenly "broke down."

"Every time I ask it to use its agentic capabilities, it gives me some vague excuse for why it can't."

This transformation from "all-powerful assistant" to "constant deflection" has left many power users confused and frustrated. What's really going on behind the scenes? Is this product degradation, or an industry-wide trend?

What Are Comet's Agent Capabilities

For those unfamiliar, here's some background. Comet is Perplexity's AI browser, and one of its core selling points is its agentic capabilities—meaning the AI can not only answer questions but autonomously operate web pages like a human: clicking buttons, filling out forms, scrolling through pages, submitting information, and so on.

This type of "AI-automated browser operation" capability represents an important direction for AI products over the past two years. From a technical implementation perspective, agentic capabilities require the AI system to possess multi-layered composite abilities: first, it needs computer vision or DOM parsing to understand webpage structure, identifying the position and function of interactive elements like buttons, input fields, and dropdown menus; second, it needs natural language understanding to parse users' high-level instructions and decompose them into specific operational steps; finally, it requires reinforcement learning-like feedback mechanisms to dynamically adjust subsequent actions based on operational results. This is fundamentally different from traditional "Q&A-style" AI assistants—an agentic AI needs to make multi-step autonomous decisions in an open environment.

From an industry landscape perspective, this direction has become a hotly contested battleground. OpenAI launched Operator in early 2025, allowing users to complete web operations through natural language instructions; Anthropic's Computer Use feature lets Claude directly control desktop environments; Google's Project Mariner focuses on agent operations within the Chrome browser; and Microsoft integrates office scenarios through Copilot Actions. Additionally, numerous startups like Adept AI and Rabbit R1 are exploring different paths. Comet is one of the representative products in this wave, attempting to seamlessly embed agentic capabilities into users' daily browsing experience through an "AI-native browser" product form.

Precisely because of this, when this core capability shows obvious degradation, user disappointment is entirely understandable.

Four Possible Causes Analyzed

The original poster proposed several hypotheses. Let's analyze each one in the context of the current industry landscape.

Websites Deploying Anti-AI Agent Mechanisms

The user specifically mentioned feeling noticeable resistance when filling out survey questionnaires. This speculation is quite plausible. As AI agent tools become more widespread, an increasing number of websites are deploying anti-automation detection mechanisms—including stricter CAPTCHAs, behavioral analysis, mouse trajectory detection, and more.

It's worth noting that the sophistication of modern anti-bot systems far exceeds what ordinary users might imagine. Taking mainstream solutions like Cloudflare Turnstile, PerimeterX, and DataDome as examples, they not only detect request frequency but also deeply analyze the naturalness of mouse movement trajectories (real human mouse paths exhibit tiny jitters and acceleration changes), keyboard input timing distribution, browser fingerprints (including WebGL rendering results, Canvas fingerprints, installed font lists, screen resolution, and dozens of other dimensions), and the integrity of the JavaScript execution environment. Some advanced systems even deploy machine learning models that identify anomalous patterns by comparing against massive baselines of real user behavior. This means that even if an AI agent can simulate the surface behavior of human operations (such as moving the mouse and typing character by character), it may still expose its "non-human" identity through deeper behavioral characteristics.

For survey sites, social platforms, and other websites that particularly care about data authenticity, identifying and blocking "non-human operations" is a hard requirement. Platforms like Facebook have historically been highly sensitive to automated behavior—their anti-scraping and anti-bot systems have been iterated over more than a decade and are quite mature, capable of detecting subtle characteristics of automation tools. Therefore, Comet hitting walls on these websites is very likely directly related to target websites actively upgrading their defenses. An ongoing technical arms race is forming between AI agents and anti-automation systems.

Vendors Tightening Capabilities to Avoid Liability

This is an easily overlooked but very real factor. When an AI agent "acts on behalf of the user," if something goes wrong—filling in the wrong form, submitting incorrect compliance training answers, or making inappropriate actions on a trading platform—the attribution of responsibility becomes very ambiguous.

For a commercial company like Perplexity, legal and compliance risks must be considered. Currently, there is no mature legal framework globally that specifically addresses liability attribution for AI agent behavior. While the EU's AI Act has established a risk classification system for AI systems, it lacks specific provisions for responsibility allocation when AI agents "act on behalf of users." In the United States, the FTC (Federal Trade Commission) has begun paying attention to consumer rights issues that AI agents might cause. The core dispute is: when an AI agent's actions lead to financial losses or legal consequences, should responsibility be attributed to the AI developer, the user who gave the instruction, or the third-party platform being operated on? This legal uncertainty itself constitutes enormous commercial risk.

Having AI complete "compliance training" on behalf of users is inherently controversial: if the training is designed to ensure users truly master certain knowledge (such as safety regulations or professional ethics), having AI do it obviously defeats the purpose of the training and could even expose employers to regulatory penalties. Vendors proactively restricting such scenarios is both risk avoidance and possibly a response to pressure from enterprise clients or regulators. Until the liability framework is clarified, "better to do less than to make mistakes" may be the most commercially rational choice.

Underlying Model Capability or Policy Adjustments

The user also speculated whether "the model itself has gotten worse." This suspicion of "stealth model downgrade" is quite common in the AI community—many users have complained about certain AI products "getting dumber" after updates.

Such discussions have a long history in the AI community. A 2023 Stanford University study attracted widespread attention, claiming that GPT-4's mathematical reasoning ability showed significant decline over several months (though OpenAI disputed this, arguing it was a benchmarking methodology issue). Possible mechanisms for such "implicit degradation" include several: vendors routing some requests to lighter models with fewer parameters to control ever-growing inference computation costs; continuous training through RLHF (Reinforcement Learning from Human Feedback) making models more conservative on safety, with the side effect of reducing task execution "decisiveness"; adjusting temperature parameters or top-p sampling strategies to reduce output uncertainty; and most directly—modifying system prompts to add more conditional instructions for refusing execution.

That said, these perceptions are often difficult to objectively verify. For users, these backend changes are completely opaque—they can only perceive differences at the results level. Compared to genuine decline in model capability, what's more likely is a tightening of safety policies—meaning the model is instructed to be more inclined to refuse or respond with "I cannot complete this operation" when facing agentic tasks, rather than actual degradation of underlying reasoning and operational capabilities. In essence, this is "artificial limitation" rather than "capability decline."

Multiple Factors Compounding

The actual situation is most likely a combination of multiple factors. Policy adjustments during product iteration, cost optimization, compliance considerations, and changes in the external website environment collectively contribute to the noticeable decline in user experience. These factors may also have mutually reinforcing relationships—for example, when anti-automation detection causes operation failure rates to rise, vendors may choose to proactively restrict these scenarios to prevent users from encountering even worse experiences (such as operations failing midway or submitting incorrect data), further narrowing the range of available features.

The Structural Dilemma Facing AI Agent Products

This individual case actually reflects the structural contradictions currently facing AI agent products.

On one hand, "letting AI autonomously operate web pages" is an extremely attractive product vision representing the future direction of productivity tools. On the other hand, this path is fraught with real-world obstacles:

• Technical arms race: A "cat and mouse game" is forming between websites and AI agents, with anti-automation detection continuously upgrading. Such arms races are nothing new in internet history—from early search engine crawlers and robots.txt protocols, to ad-blocking plugins versus anti-blocking technology, to today's AI agents versus anti-bot systems, each round of technical confrontation drives spiral improvement on both sides, ultimately forming a new equilibrium.
• Liability boundaries: The legal and ethical responsibilities arising from AI acting on behalf of humans still lack a clear definitional framework. This involves not only direct economic loss compensation but also privacy compliance (AI agents may encounter users' sensitive information during operations), contract validity (whether AI clicking "agree" on behalf of users has legal binding force), and other complex issues.
• Reliability vs. safety trade-off: Vendors must either make AI more "bold" and accept the risk of errors, or make it more "cautious" at the expense of utility. This is a classic product design dilemma—in the absence of robust error recovery mechanisms and liability insurance systems, commercial rationality almost inevitably pushes vendors toward the more conservative route.

For products like Comet, scenarios that were initially opened up to showcase capabilities being gradually tightened as user scale expands and risks become exposed is almost inevitable. This also explains why early-adopting power users feel the strongest sense of "gap"—their usage habits were built on the capability boundaries of the product's "aggressive phase," while the product is transitioning to a "stable phase."

Practical Advice for Users

If you're also a user of Comet or similar AI agent products, consider the following perspectives:

First, understand that the product's capability boundaries are dynamically changing. Features that work today may be restricted tomorrow—this is the norm for current AI agent products, and critical workflows shouldn't be entirely dependent on them. It's advisable to maintain manual operation as a backup for important tasks, positioning AI agents as "accelerators" rather than "replacements."

Second, for serious scenarios involving liability such as compliance training and formal form submissions, even if technically feasible, having AI do the work carries risks—use with caution. Particularly in workplace environments, being caught using AI to complete compliance training could result in disciplinary action or even legal consequences.

Finally, pay attention to official update notes and community feedback. Many capability changes are not proactively announced by vendors, and exchanging information through communities (such as relevant Reddit boards) is often the fastest way to learn the real situation. Also keep an eye on competitor developments—if multiple AI agent products simultaneously tighten certain types of functionality, it usually indicates an industry-level compliance adjustment rather than a single product's issue.

Comet's "becoming timid" is less a single product's failure and more a growing pain that the entire AI agent sector must inevitably experience on its path to maturity. Finding the balance between utility, safety, and compliance will be a long-term challenge for all AI agent products. This process may require technological breakthroughs (such as more reliable error recovery mechanisms), institutional innovation (such as AI operation insurance and clear liability-sharing frameworks), and the establishment of industry standards (such as behavioral protocol specifications for AI agents) working together to drive progress.

Key Takeaways