Personal AI Agents: Best AI That Can Actually Do Things for You

Updated 2 November 2026. Personal AI is moving from answering questions to actually doing the errand. The useful test is no longer “does it sound clever?” but “did it finish the job without quietly changing my constraints?”

Pick your fighter: what do you actually want done?

  • Research: use an agent that can gather evidence, compare sources and show its work.
  • Shopping: prioritise constraint-following, transparent recommendations and safe checkout.
  • Travel & booking: look for live availability, confirmation and a clear approval step before spending.
  • Phone calls: this is becoming a category of its own. Agents can increasingly call restaurants, shops and service businesses to complete the awkward last mile. Great if the task is routine; bad fit if a misunderstanding can create a costly commitment. Insist on a transcript or action summary, confirmation of what was agreed and approval before payment or a non-refundable booking.
  • Computer use: potentially the broadest category, but permissions and recoverability matter most.

Claude Duet vs OpenAI Dots vs Meta Muse: proactive agents compared

New on 21 October: Claude Duet joined the proactive-agent race. It offers morning briefs, connected-app monitoring and background work for professionals, but is currently a US Pro/Max waitlist research preview. Its background checks consume plan usage, and Anthropic says personal errands are still weaker. Official launch details. For the three-way comparison, pick your fighter here; for measuring whether an agent actually completed a job, use our agent evaluation checklist.

Consumer agents are becoming action layers rather than prettier chat windows. Meta’s Muse has pushed the category into the mainstream with research, forms, purchases, reservations and connected services. OpenAI’s Dots, announced on 29 September, pushes the same idea toward persistent, always-on work across connected apps. The useful comparison is no longer which one writes the nicest answer. It is which agent can finish your task, preserve your constraints and show you exactly what it did.

  • Pick Muse if: you care most about consumer errands, shopping and Meta-connected experiences.
  • Pick Dots if: your priority is persistent work across business apps and ongoing projects.
  • Skip both for high-stakes autonomy: if a mistake can move money, expose sensitive data or change an important account, keep a human approval step.

Our take: this category is finally interesting because the products are starting to act, not merely chat. That also makes permissions, receipts and revocation part of the product review — not boring security footnotes.

The bit the demos skip: can the agent be trusted?

The more useful agents become, the more consequential their mistakes become. On 24 September, Australia opened an investigation after an OpenAI agent was reported to have gained unauthorised access to a Medicare healthcare database. Separately, banks have warned that shopping agents create new fraud, privacy and payment risks, while Amazon has blocked Meta’s Muse from shopping on its marketplace. The practical lesson is simple: capability without permission boundaries is not a feature.

  • Prefer narrow permissions: give an agent access to the smallest set of accounts and actions it needs.
  • Require approval before money moves: purchases, bookings and account changes should have a visible checkpoint.
  • Demand receipts: transcripts, confirmations, source links and action logs make mistakes easier to spot.
  • Check revocation: you should know how to disconnect the agent and invalidate its access.
  • On a Mac, treat Full Disk Access as high-risk: Apple said on 2 October that increasingly capable AI agents have raised the risks of this permission, which can expose messages, email, files and browsing data. Prefer narrower access where the product supports it. On 8 October, the UK ICO opened a call for evidence on agentic AI risks, including data security, transparency and accountability. Those belong in every agent buying checklist.

How we judge them

  • Completion: did the task actually finish?
  • Constraint fidelity: did the agent preserve dates, budget, location and preferences?
  • Approval: did it ask before consequential actions?
  • Evidence: can you see what it researched, clicked, booked or bought?
  • Recovery: what happens when a site, payment or phone call behaves unexpectedly?
  • Time saved: checking the agent should be easier than doing the whole task yourself. Run the same task twice with different constraints; if the outcome changes unpredictably, you are still the quality-control department.

Great if…

You have repetitive research, comparison-heavy decisions, admin or multi-step digital errands where the cost of checking the result is lower than doing everything manually.

Bad fit if…

A plausible mistake is expensive, permissions are unclear, or the agent cannot prove that the task was completed. Shopping is a good example: banks are already warning that agentic commerce can introduce fraud, privacy and recourse problems when an AI starts acting between the customer and payment system.

Our take

The hierarchy that matters is assistant → agent → trusted agent. An assistant tells you what to do. An agent tries to do it. A trusted agent has repeatedly shown that it can finish correctly, transparently and within boundaries. Most products are still earning that last label. A useful next test is repeatability: can the agent finish the same task under changed conditions, and can you inspect its actions when it fails? A polished demo is not an evaluation.

Work agents: Gemini joins the fight

On 8 October 2026, Google introduced a Gemini agent for work. It can plan tasks across Workspace. Coworker agents get separate identities and limited access. Availability is currently a private enterprise preview. Great if your team works in Gmail, Drive and Docs; bad fit if you need a consumer shopping agent today. Compare Gemini. Source: Google Cloud, 8 October.

Pick your fighter for work: Gemini is worth watching for Workspace-heavy teams; Dots targets persistent cross-app work; a conventional assistant is enough for drafting and answers. Trust Check: ask whether the agent has its own identity, whether admins can restrict data access, and whether every action appears in an audit log.

AI agent identity, passports, permissions and shadow AI

14 October 2026: Sophos announced general availability of AI Defense, a security product designed to discover unapproved AI tools and inspect activity across devices, browsers, networks and identities. This is an enterprise governance tool, not a consumer assistant. For a useful pilot, ask whether it can distinguish personal from corporate AI accounts, flag sensitive uploads, identify connected agents and revoke access. Official Sophos announcement.

Agent passport comparison: ServiceNow, Okta and Cloudflare

Workday announced an agreement to acquire C1.ai on 13 October 2026. On 15 October, ServiceNow announced AI Passport, with availability expected in Q4 2026. It proposes signed, scoped permissions for first- and third-party agents, with revocation and audit trails. This is a roadmap announcement, not proof that the controls are live or independently tested. Compare it with Okta for AI Agents, which already offers agent identity registration and lifecycle governance, and Cloudflare Mesh, which focuses on private connectivity and access policies. They solve overlapping but different problems: agent ownership, agent authorisation and network isolation. Pick ServiceNow if your organisation runs its workflows there and can wait for the Q4 launch; pick Okta if cross-vendor identity lifecycle is the urgent problem; pick Cloudflare Mesh if the immediate priority is controlling which private services an agent can reach. For every vendor, request a live demonstration of a denied permission, an audit log tied to a named owner and immediate session revocation. The deal has not closed. Agent identities, narrow permissions and revocation are now practical buying criteria, not optional IT extras. Workday announcement.

Trust Check: Include runtime tool-call controls, independent evaluation, clear human approval points and an auditable record of every action. On 16 October, Reuters reported Tencent acknowledging unintended browsing by agents used in third-party model evaluations. That makes auditability and network limits a practical requirement, not a theoretical one. Ask six things: does the agent have its own identity? Can access be restricted to one task? Who approves changes or payments? Can you inspect its actions? Does access expire? Can you revoke it immediately?

Great if: your team is deploying agents into sensitive systems. Bad fit if: you just need drafting or everyday search. Ask for a failed-permission demo, not another glossy video.

AI shopping agents: discovery versus authorised checkout

Pick your fighter: research, tracking or buying?

Research: ChatGPT or Perplexity for comparing options and checking sources. Price alerts: Gemini if Google Universal Cart is available in your region. Buying: only delegate checkout when the agent shows the exact seller, total cost, approval and cancellation path. A slick recommendation is not a receipt. Compare ChatGPT, Gemini and our side-by-side guide.

New for October 2026: Google Universal Cart launched in the US on 15 October, bringing multi-store product saving and price alerts. Shopping agents increasingly differ on merchant access and checkout approvals. American Express published its Intent Intelligence framework on 13 October 2026 to preserve a shopper’s authorised purchase instructions across AI-assisted transactions. It is a proposed technical approach, not proof of a general consumer rollout. Shopify powers product discovery and checkout for participating merchants across several AI channels; it is not itself a consumer shopping assistant. On 13 October, Shopify opened Shop MCP tools for agents to search its catalogue, build carts and use Shop Pay checkout. Availability and user approval still depend on the specific integration. On 16 October, Shopify added a buyer-consent extension to its cart and checkout tools, allowing supported agents to carry a shopper’s privacy and marketing choices into checkout. That is separate from authorising a purchase. TikTok announced Shopping Assistant and Buy Direct on 5 October; these are platform-led shopping experiences, not independent cross-store research tools. We compare documented capabilities, merchant restrictions and checkout steps — not glossy demos. For price alerts start with Gemini; for broad product research start with ChatGPT. Before trusting any shopping agent, check five things: seller identity, the full delivered price, returns, explicit approval and whether you can cancel the instruction. If the assistant can only recommend products, that is useful research, not autonomous checkout. Check shipping, returns and seller identity before paying.

Shopping agents · 21 October 2026

WebMCP: the difference between clicking blindly and using a shop’s actual tools

The useful bit: AI shopping agents often try to read buttons and click through storefronts. WebMCP lets a participating website expose structured actions such as product search and adding items to a basket. The agent can ask the shop directly instead of guessing where to click.

What’s new: Cloudflare and Commerce.com announced on 21 October that their integration brings WebMCP tools to almost 100,000 Commerce.com storefront domains, with merchant opt-outs. The feature is enabled for eligible stores, but WebMCP is still an experimental browser standard; this is not proof that every AI assistant can shop at every participating store today.

RouteGreat if…Watch out for
Browser clicksA shop has no agent toolsCookie banners, bot checks and wrong variants
WebMCPA supported agent is on an enabled storefrontEarly-preview browser support and merchant opt-outs
Server-side MCPAn agent connects to a shop’s service directlyIntegration and authorisation still vary

Pick your fighter: research, basket or checkout?

Research across shops: start with ChatGPT or Gemini, then verify prices at the seller. Use an agent on a supported shop: ask whether its browser supports WebMCP and whether that merchant enables it. Ready to pay: check the final basket, delivery, returns and seller identity yourself. In the Commerce announcement, checkout stays with the merchant and the human shopper.

Skip it for now if you need guaranteed unattended purchasing across arbitrary websites. WebMCP improves the plumbing, not permission, availability or judgement. See our assistant comparison and AI directory for adjacent tools.

Get Worth a Tab shopping updates · Do the shopping-agent research for me · Get a tool listed

Shopping agents · updated 24 October 2026

Walmart’s AI Agent Profiles: who gets to spend your money?

New on 21 October: Walmart announced Agent Profiles, an Agent Gateway and new rules for AI shopping assistants. Agent Profiles are scheduled to roll out within the following month. The Gateway is starting with an invitation-only developer beta. Do not mistake a product announcement for general availability.

What changes: shoppers should be able to authorise each approved agent separately, see its purchases apart from their own, require approval for every order or set spending limits, and revoke access. Walmart says agents cannot access card details or change payment methods; its returns process applies if an agent orders the wrong thing. Those are documented policies, not independent reliability tests.

Pick your fighter: research, price alerts or buying?

Walmart vs Amazon vs Shopify: Walmart is preparing a permissioned gateway for approved agents; Amazon restricts some unauthorised third-party agents; Shopify’s Shop MCP gives compatible agents catalogue, basket and Shop Pay checkout tools. These are different routes, with different availability. See the 21 October retailer comparison.

The five-minute test: ask an available assistant to find a specific product, size and quantity under a strict delivered-price cap, then stop before purchase. Check the seller, substitutions, total, approval and cancellation path. If it cannot reliably obey those instructions in rehearsal, do not grant an unattended spending limit. See our agent evaluation checklist.

Great if: you want repeat grocery purchases with visible controls. Skip it for: expensive, non-refundable or size-sensitive purchases until the exact approved agent and checkout flow are proven. Protected card data does not guarantee the agent will choose the right product.

Request Worth a Tab shopping updates →Do the research for me →Get listed →

Sources: Walmart (21 October), Shopify (13 October), TechCrunch (21 October). Based on published capabilities; no hands-on testing of unreleased features claimed.

Start with the AI Tools & Agents Directory, compare the major assistants in our comparison hub, or read the individual ChatGPT, Claude, Gemini and Perplexity reviews.

FAQ

What is a personal AI agent?

An AI system designed to carry out tasks on your behalf, often across several steps or tools rather than only answering a question.

How is an AI agent different from a chatbot?

The useful distinction is behaviour: a chatbot primarily responds; an agent can plan and take actions toward an outcome.

Should I give an AI agent access to my accounts?

Start with narrow permissions and low-risk tasks. Understand what the provider can access, what actions require approval and how you can revoke access before trusting consequential workflows.

Updated 2 November 2026 · Independent browser-agent comparison

Best AI browser agents: Claude in Chrome vs Comet vs Gemini vs ChatGPT Work

The question is no longer “can it summarise a website?” It is “can it finish a web task without clicking the wrong button, sharing my data or spending my money?” These four products solve different parts of that problem. Our recommendations are based on vendor documentation, not a hands-on performance benchmark.

Pick your fighter: what are you asking it to do?

ToolWhere it worksImportant boundary
Claude in ChromeChrome desktop extension; paid plansBroad browser permissions; approval modes; rollout
CometMac, Windows, iOS, Android browserFree download does not imply unlimited agent actions
Gemini in ChromeEligible Chrome accountsAuto browse plan/region restrictions, US-first rollout
ChatGPT WorkSeparate cloud browser, supported paid plansSeparate sign-in; some websites and transactions unsupported

Claude in Chrome: for supervised multi-tab work

Anthropic’s Chrome extension can read, click, type, navigate and work across grouped tabs. It also documents console-log access and scheduled browser tasks. Access is tied to eligible paid plans, and some capabilities depend on rollout and enterprise settings. Great if: you want to watch an assistant operate your existing Chrome tabs or verify a web app. Skip if: you need a mobile browser replacement or cannot grant the extension’s requested permissions. Our Claude profile · Official guide ↗.

Perplexity Comet: for an AI-first browser

Comet pairs browsing with Perplexity research, summaries and supported actions. Perplexity lists Mac, Windows, iOS and Android; its browser is free to download, but heavier agent features can be plan-dependent. Great if: research and browsing happen together on several devices. Bad fit if: you assume every shopping or calendar workflow is included without limits. Our Perplexity profile · Official product page ↗ · Comet and background-assistant announcement ↗.

Gemini in Chrome: for eligible Google-first users

Google describes auto browse for multi-step errands, forms and shopping, with sensitive steps returned to the user. Its published eligibility guidance lists auto browse for Google AI Pro and Ultra subscribers in the US; broader availability must be checked for your account. Great if: your work is already in Chrome and Google services. Skip if: you require guaranteed non-US access today. Our Gemini profile · Google’s eligibility guidance ↗ · 30 July Spark update ↗.

ChatGPT Work: for delegated browser workflows

OpenAI’s cloud browser operates on a separate computer with its own sessions, so it does not automatically inherit your laptop’s logins. It can work on supported websites and pause for sign-in or confirmation; availability depends on paid plan, region and rollout. Great if: you want to hand off a multi-step research or admin task. Bad fit if: you expect unrestricted access to every site or transaction. Our ChatGPT profile · Cloud browser guide ↗.

Still comparing ChatGPT Atlas with Comet?

Check the publication date. Atlas was discontinued on 9 August 2026; OpenAI moved its browser-agent efforts into ChatGPT and Codex. A ranking that still recommends Atlas is out of date. OpenAI’s migration notice ↗.

The 10-minute test: does it finish without crossing a line?

  1. Ask for three refundable hotels on exact dates, under a strict total budget. Say do not book.
  2. Change the dates mid-task. Does the assistant keep the new constraints?
  3. Verify taxes, availability, refund rules and source links manually.
  4. Ask it to proceed towards checkout. Does it stop before an irreversible action?
  5. Check its action history and whether you can revoke access or recover from a mistake.

Our verdict: a polished summary is not a completed task. Pass only if every constraint held, the evidence is verifiable and no unapproved action occurred. For a full testing framework, see AI agent evaluation vs monitoring.

FAQ: what buyers actually ask

Which AI browser agent is free?

Comet is free to download, but advanced agent use may require a paid Perplexity plan. Other browser-agent features have paid-plan and regional restrictions. Compare task allowances, not just the download price.

Can AI browser agents safely make purchases?

Some can operate shopping flows, but website support and approvals vary. Use explicit confirmation before payments, messages, account changes or accepting terms. Never give an agent blanket spending permission.

Is ChatGPT Atlas still supported?

No. OpenAI says Atlas stopped working on 9 August 2026. Look at supported ChatGPT browser features instead.

Worth a Tab: the AI tools worth opening

Get independent updates when we publish. Need a shortlist for your particular workflow? Ask us; no paid rankings or pretend testing.

Browse the AI directory, assistant comparisons and related guides.

Research checked 2 November 2026 against official product documentation. These are evidence-based starting points, not independent hands-on performance rankings. Feature availability and pricing can change.

Scroll to Top