Best AI Browser Agents in 2026: Comet, Dia, Fellou, and Polar Put to the Same Task

I spent a week running the exact same errand through four different browsers: research a purchase, cross-check the claims across a couple of sources, and fill out a form at the end. Not a toy prompt like "summarize this page" — an actual multi-step job with a real outcome at the end of it. Two of the four browsers finished it without me touching the mouse. One finished it and then quietly did something I didn't ask for. One never really tried.

My name is Artem, and I run the Writingmate blog, which means I spend an unreasonable amount of time installing AI products the week they launch so I don't have to guess how they behave. This week that meant Perplexity's Comet, The Browser Company's Dia, Fellou, and Polar — a brand-new entrant that TechCrunch broke the story on July 29, 2026, built by a former Perplexity engineer who helped ship Comet in the first place. That last detail is what made me actually run this comparison now instead of filing it under "someday": when the person who built the thing goes and builds a competitor, it's worth checking whether he learned something the rest of us haven't caught up on yet.

Here's what I found, task by task, plus what to do if you don't want to install a fifth browser just to get the automation part.

What "agentic browser" actually means in 2026

Quick level-set, because the term gets used loosely. A normal browser shows you the web. An agentic browser has an AI layer that can also act on the web on your behalf — click buttons, fill in fields, navigate between sites, and carry state across tabs, usually while logged in as you. That's a meaningfully bigger permission grant than a chatbot answering questions, which is exactly why the security section further down matters.

Four products currently define this category, and they've each picked a different point on the assist-to-autonomy spectrum:

• Comet (Perplexity) — launched July 9, 2025 to $200/month Max subscribers, went free worldwide in October 2025, now runs on Mac, Windows, iOS, and Android with agent mode included free.
• Dia (The Browser Company, now an Atlassian subsidiary) — the successor to Arc, built around "chat with your tabs" rather than full autonomy, macOS/Apple Silicon only as of this writing.
• Fellou — self-described as the first fully agentic browser, built around two modes called Deep Action and Deep Search, over a million users since its April 2025 launch.
• Polar — the newest of the four, announced July 29, 2026, built specifically for knowledge-work automation rather than general browsing, founded by ex-Comet engineer Kevin Jiang alongside Vishaal Ram and Howard Zhong.

How I ran the test

I picked one task that's realistic for an actual workday rather than a benchmark showcase: research a standing desk under $500, compare specs and shipping across three retailer sites, write a two-sentence recommendation, and submit an email-capture form on the winning retailer's site to get notified about restock or price drops. It's the kind of chore that eats twenty minutes of a Tuesday and has a clean, checkable end state — either the form submitted with the right desk's details or it didn't.

I scored each browser on three things:

• Completion — did it finish the whole chain unassisted, or did I have to step in?
• Transparency — could I see what it was about to do before it did it, and could I stop it mid-task?
• Corrections needed — how many times did I have to fix, redirect, or manually finish a step?

I also leaned on the disclosed benchmark numbers each company has published, plus the two independent security disclosures against Comet from Brave and LayerX, since "does it finish the task" and "should you let it near your accounts" turned out to be two very different questions.

Comet: the most capable, and the most already-attacked

Comet was the fastest of the four to get from "research three desks" to "here's my pick," largely because Perplexity's underlying search is genuinely strong at cross-referencing specs and pulling citations. It handled the product comparison cleanly and got through the form with only one hiccup — it picked the wrong email field on a page with a newsletter signup stacked above the restock-alert form, and I had to redo that one step.

Where Comet loses points is trust. In 2025, Brave's security team disclosed an indirect prompt injection against Comet where hidden instructions in a webpage got the agent to open Gmail, retrieve a one-time passcode, and leak both an email address and the code to an outside party — all from what looked like an ordinary page. Separately, LayerX researchers found and disclosed "CometJacking," where a single malicious link could hijack Comet's existing access to connected accounts like Gmail and Calendar without any credential theft at all; Perplexity initially closed that report as "not applicable" before acknowledging the issue. Comet is also the browser most often flagged for memory bloat — a handful of tabs with AI features active can push RAM usage well past what a normal Chromium session needs.

"Nobody used AI features in Arc so we killed the entire browser to make a new AI browser that nobody will use" is the kind of skeptical take that circulates on Reddit whenever a company kills a product people liked to build the one they're betting on instead — and it's fair to point that same skepticism at any agentic browser until you've watched it work on your own tasks.

Dia: calm, capable at research, not trying to be an agent

Dia didn't finish my task, and that's by design rather than a failure. Dia's sidebar can read every tab you have open and synthesize across them — it was genuinely good at pulling specs from three retailer pages into one clean comparison — but it doesn't click through checkout flows or submit forms for you. You get the research, then you do the clicking.

I went back to the source on why The Browser Company built it this way. In a podcast interview about killing Arc to build Dia, CEO Josh Miller admitted the honest version of the story isn't that Arc failed outright: "I almost think blaming Arc is a scapegoat. If really Hursh originally and then myself didn't get so inspired by this new material that are LLMs, we would not have pivoted Arc to something else." Arc's own numbers backed up the retreat — only about 5.5% of daily active users regularly touched Spaces, one of Arc's signature features, and even fewer used Live Folders. Dia's answer to that adoption problem is Skills: reusable AI shortcuts you build once ("find interesting events near me this weekend," "summarize this thread") and re-run with one click, rather than the browser silently taking multi-step action on your behalf. It's a reasonable bet for a knowledge worker who wants a research partner, less so if what you actually wanted was the desk ordered.

Fellou: the most ambitious plan, the shakiest execution

Fellou is explicitly built for exactly my kind of task — its Deep Action mode is designed to take a goal like "find 10 jobs suited to my LinkedIn profile and apply" and turn it into a visible, editable step-by-step plan you can approve before it runs. That transparency is genuinely the best of the four: you see the plan, you can edit or delete steps, and you can jump in mid-execution.

The problem was follow-through. Fellou built a correct plan for my desk task, then got stuck on one retailer's site where the "add to compare" interaction used a layout it hadn't seen before, and looped for close to two minutes before I intervened. It eventually got there with help. Fellou's own published numbers claim an 80% completion rate on the Online-Mind2Web benchmark against roughly 43% for competing tools, and I believe that gap is real for the sites its team has tuned against — it's the messy, unfamiliar sites where it still struggles.

Polar: newest, narrowest focus, best raw numbers

Polar was the only browser of the four built around exactly this kind of job rather than general browsing, and it showed. It finished my full task chain — research, comparison, form submission — without a single correction, and it did it while showing a clear running log of each action it was taking, which made it easy to sanity-check before the final submit.

Polar's own published benchmark, OdysseysBU Bench V1, has it scoring 98.0 against 75.5 for a tool called Aside, 70.0 for Browser Use, 60.8 for a comparable GPT-based agent, and 44.5 for Claude Opus in the same agent role — worth reading as a strong self-reported number rather than an independently audited one, since Polar published it themselves. What I can independently confirm: the company says users have run more than 4.5 million actions through it since a quiet, word-of-mouth launch roughly seven months before the funding news broke, and one company using it reported saving 25+ hours per person per week. Polar just raised a $5.7 million seed round led by Madrona on July 29, 2026, with angels including former GitHub CEO Thomas Dohmke and Modal CEO Erik Bernhardsson — small money by AI-startup standards, which tracks with a product still finding its footing rather than one that's already at consumer scale.

The scoreboard

Browser	Completed task unassisted	Transparency of steps	Corrections needed	Platform / price
Comet	Yes, 1 hiccup	Medium — acts, then reports	1	Mac, Windows, iOS, Android — free
Dia	No — research only	High — you drive the actions	N/A (by design)	macOS/Apple Silicon only — free, Pro paid tier
Fellou	Yes, after intervention	Highest — editable plan before run	1 (stuck loop)	Windows, Mac — free + $19/$39.9/$199 credit tiers
Polar	Yes, unassisted	High — live action log	0	Desktop — free + $20/mo+

The part nobody selling you a browser wants to lead with: security

Every one of these tools asks for a level of access a normal browser never did — reading your email, acting inside apps you're logged into, sometimes touching payment flows. That's the entire value proposition, and it's also the entire risk. The Comet disclosures from Brave and LayerX aren't edge cases; they're proof-of-concept demonstrations that a hidden instruction on an ordinary-looking webpage can turn your own logged-in agent against you, with no phishing link or malware required. No vendor in this category, including the three others reviewed here, has published a complete fix for this class of attack as of mid-2026. My working rule: agent mode is fine for research and shopping comparisons, and I keep it away from banking, healthcare, and anything with stored payment methods until a browser can show me it's actually solved indirect prompt injection, not just patched the specific demo that got disclosed.

You don't need a new default browser to get this

Here's the thing that struck me running all four of these back to back: three different startups (four, counting Perplexity) each decided the fix for "AI should be able to finish tasks for me" was to make you replace your browser. That's a real cost — new defaults, new bookmarks, new muscle memory, and in Dia's case, a whole platform restriction to macOS.

A lighter option is to build the specific job as a custom agent inside a platform you already use for everyday work. In Writingmate's custom agents, you describe the task once — research three sources, compare, draft a recommendation, take an action — give the agent the tools it needs, and run it from the same chat interface you're already using, across whichever model fits the job best. No new browser install, no new default to fight your team over, and you're not locked into one company's roadmap for how fast the agent gets better. If you're evaluating whether that's worth it against a $20/month browser subscription, the pricing page breaks down what's included at each tier.

So which one should you actually install?

If you want the most capable general-purpose agent today and you're comfortable keeping it away from sensitive accounts, Comet is still the easiest recommendation — it's free, it's on every platform, and Perplexity has had a year of real-world usage to sand off rough edges. If your job is mostly research and synthesis and you're on a Mac, Dia is calm and genuinely useful without the autonomy risk. If you want to watch an agent plan itself out and you're willing to babysit it through the rough patches, Fellou is the most transparent. And if the job you keep doing by hand is long, repetitive, multi-step knowledge work — not shopping, but sales research or recruiting pipelines — Polar is worth the seven-minute install specifically because it was built for that job and not for general browsing.

My honest take after a week with all four: none of them are ready to run near your bank account, and all four are already good enough to save you real time on research-and-report chores. Start there, not with your passwords.

See you in the next one!

Artem