Skip to main contentSkip to footer

    Top Rated & Verified

    Top Clutch App Development Company Black Owned United StatesTop Clutch Java Developers France 2026Top Clutch Service Line Blind Company Black Owned 2026Top Clutch App Development Company Minority Owned 2026Top Clutch Web Developers Black Owned 2026Top Clutch App Development Company Black Owned 2026Top Clutch Flutter Developers France 2026Top Clutch Health & Wellness App Developers France 2026Top Clutch Swift Company France 2026Top Clutch Machine Learning Company France 2026Top Clutch Chatbot Company France 2026Top Clutch Artificial Intelligence Company France 2026Top Clutch App Development Company Minority Owned Los Angeles
    Back to Blog
    AI Agents
    November 15, 2026
    29 min read

    Best AI Agents in 2026:10 General Agents Ranked on What You Can Check

    Ten general agents scored on five checks that you can repeat: price, free start, approval step, prompt injection defense and training default. I read every page on November 15, 2026. My own company is not in the scored list.

    A city skyline at dusk under a network of glowing teal circles and lines, in purple and pink light
    5 of 10
    general agents document an approval step before sensitive actions that is on by default
    Frenchy Digital review of ten vendor pages, November 15, 2026
    6 of 10
    vendors name prompt injection and describe a defense against it
    Frenchy Digital review of ten vendor pages, November 15, 2026
    2 of 10
    vendors say that they do not train models on your data by default
    Microsoft Learn and Opera Neon pages, checked November 15, 2026
    $0
    the price of ChatGPT Work on the ChatGPT Free plan
    ChatGPT pricing page, checked November 15, 2026

    Key Takeaways

    • I scored ten general AI agents on five checks from public pages, read on November 15, 2026. Microsoft Copilot Cowork and ChatGPT Work tie at 9 of 10. Alexa+ is last at 3.
    • Five of the ten document an approval step before sensitive actions that is on by default. Their lists agree: purchases, payments, sent messages, deletions and sign-ins.
    • Only two vendors, Microsoft and Opera, say that they do not train models on your data by default. Four others give you a setting to opt out.
    • ChatGPT Work is the only agent in the top five with a free plan. The paid entry price for most consumer agents is about $20 a month.
    • No vendor claims to stop prompt injection fully. Perplexity says that it remains unsolved across the industry, so keep the approval step on.
    • Frenchy Digital is not in the scored list. A custom agent from us costs $10,000 to start and $2,500 a month, so an agent from this list costs less for most general tasks.

    Ten agents, five checks

    On November 15, 2026, I read the public pages of ten general AI agents. My goal was to rank the best AI agents on facts that you can check yourself, not on a demo.

    Five of the ten say that the agent asks you before a sensitive action, such as a purchase or a sent email, by default. Six name prompt injection and describe a defense. Two say that they do not train models on your data by default.

    Those three numbers are the real story of this list. A general agent reads your email, opens web pages and uses your accounts. Before you give it that access, you want to know three things. When does it stop and ask you? What protects it from hostile text? Where does your data go?

    So I scored each agent on five checks. They are a published price, a free start, an approval step, a defense against prompt injection and the training default. Each check gives 0, 1 or 2 points, and the maximum is 10.

    I am Chris Machetto. I run Frenchy Digital, and we build custom agents for clients. That gives me a commercial interest, so I did not score my own company. It appears after the list, in a section that I label as an advertisement.

    This list covers general agents. A general agent takes a goal from a person, then does the steps in a browser, on a computer or in connected apps. For agents that run a business process, read our ranking of the best AI agents for business automation. It scores ten different products on different checks.

    What a general agent is

    A general agent takes a goal in plain language and does the steps itself. It does not wait for one instruction at a time. It returns a result: a booked table, a sorted inbox, a report or a finished file.

    The vendors describe it in similar words. OpenAI opens its guide to ChatGPT Work with one short sentence.

    ChatGPT Work is a way to delegate real work to ChatGPT.

    OpenAI, Get started with ChatGPT Work

    Meta uses a list of tasks. In its launch post for Muse, Meta says that the agent can send emails, book travel, fill out forms and negotiate for you. It continues to work after you close the app.

    The ten agents in this list work in three places. Some work on a computer in the cloud. Some work on your own computer. Some work in your browser, where you are already signed in. The place matters, because it decides what the agent can reach.

    AgentWhere it worksWhat it can reach
    Microsoft Copilot CoworkMicrosoft cloud, plus a hidden tab in Microsoft Edge on your deviceMicrosoft 365 files and mail, plugins, websites where your sign-in is active
    ChatGPT WorkA cloud container, or your computer through the desktop appYour files, plugins such as Slack and Google Drive, the browser
    Claude CoworkThe cloud, and your computer when the desktop app is openLocal files, the browser, connected tools
    Gemini SparkGoogle cloud, plus Chrome for web tasksGmail, Calendar, Docs, Sheets, Drive, websites where your sign-in is active
    OpenClawYour own hardwareChat apps such as WhatsApp and Slack, files and commands on the host
    Meta MuseA dedicated cloud computer for each userConnected services, the web through the Sentinel
    Perplexity ComputerAn isolated cloud sandbox, or your own Mac or NVIDIA hardwareGmail, Outlook, Slack, Notion and other connected apps
    ManusA cloud browser, or your own browserAccounts that you sign in to
    Opera NeonThe Opera Neon browserWebsites in the browser
    Alexa+Amazon cloud and Alexa devicesPartner services for takeout, reservations, rides and repairs

    Read the last column with care. An agent that uses your signed-in browser has the same access as you. Microsoft says it directly about Cowork: the agent has exactly the same access to sites that you have, no more and no less.

    This article is about agents for a person or a small team. For a company view, read our guide to AI agents for business first. It explains what agents do in each department and what a project costs.

    Therefore: before you compare agents, write down what each one can reach. An agent that can reach your bank, your inbox and your files needs stronger controls than an agent that writes a report.

    The valet key

    Many buyers think that the best agent is the one that can do the most without them. That belief is reasonable. The demos show an agent that books a trip from start to end while the person watches.

    The belief misses one point. An agent that can do everything alone can also make every mistake alone. The best general agent is the one that stops at the correct moment.

    Car makers solved a similar problem a long time ago with the valet key. A valet key is a car key that starts the engine and opens the doors, but not the trunk or the glove box. The valet can park the car. The valet cannot open what holds your valuables.

    Apply that idea to an agent. The agent may read, search, compare and write drafts. It needs your approval for the trunk: a payment, a sent message, a deletion or a sign-in. The approval step is the valet key.

    The model is not the key. Microsoft Copilot Cowork launched on Anthropic models, and Anthropic sells the same family of Claude models in Claude Cowork. Redmond Channel Partner reported that Cowork launched on Opus 4.8 and Sonnet 4.6. The two products have similar models but different prices, approval rules and data rules.

    If two products can use the same model, the model does not decide which one is best for you. The controls around the model decide it.

    Therefore: I ranked the agents on the controls that you can check, not on how smart they seem in a demo. A smarter model with no approval step is a valet who holds every key to the car.

    How I scored them

    I selected ten general agents that a person or a small team can start to use today. The list has agents from Microsoft, OpenAI, Anthropic and Google. It also has two consumer platforms, a research company, a browser maker, a startup and one open-source project.

    I excluded agents that only write code, and products that only run business workflows. One product, Genspark, blocked my reader on its pricing page, so I left it out. I cannot score a page that I cannot read.

    The five checks

    Each check gives 0, 1 or 2 points, so the maximum is 10. These are the rules for each check.

    • Price.: A dollar or euro price on a vendor page for a plan that includes the agent gives 2. Free software under an open-source license also gives 2. Paid plans with no figure give 1.
    • Free start.: A free plan that includes the agent gives 2. A free trial, a one-time free allowance, or a free plan without the agent gives 1.
    • Approval step.: A vendor page that says the agent asks you before sensitive actions, by default, gives 2. A control that is off by default, that covers only sign-ins, or that applies only to a related product gives 1.
    • Injection defense.: A vendor page that names prompt injection and describes a defense gives 2. A control on outside actions that does not name prompt injection gives 1.
    • Training default.: A vendor page that says it does not train models on your data by default gives 2. A documented setting to opt out, or a promise for business plans only, gives 1.

    What I excluded. I did not score accuracy, speed or return on investment. The vendor is the source of each such number on its own site, and no independent test measures these agents as the vendors sell them. I did not show app store ratings, because a rating covers the whole app, not the agent inside it.

    When and how. I read every page on November 15, 2026. A tie goes to the agent with more points on the three safety checks: approval, injection and training. If the tie remains, it goes to alphabetical order. A zero means that I did not find the fact on the pages that I read, not that the feature does not exist.

    RankAgentPriceFree startApproval stepInjection defenseTraining defaultTotal
    1Microsoft Copilot Cowork212229
    2ChatGPT Work222219
    3Claude Cowork212218
    4Gemini Spark212218
    5OpenClaw221218
    6Meta Muse122117
    7Perplexity Computer211217
    8Manus221005
    9Opera Neon200024
    10Alexa+210003

    The ten, ranked

    Each entry says who the agent suits, what I verified and what I did not find. Read the last part with the same care as the first.

    Top ranked
    1

    Microsoft Copilot Cowork

    The most documented controls, sold only with Microsoft 365 business plans

    Our score: 9 of 10

    Suits: companies that work in Microsoft 365 and have an admin who can set spending limits.

    What I verified: Microsoft made Cowork generally available on June 16, 2026, and it requires a Microsoft 365 Copilot license. The pricing page for business lists the Copilot Business add-on at $21 per user a month, billed yearly. A promotion takes it to $18 for the first year, until December 31, 2026. The add-on needs an eligible Microsoft 365 Business plan.

    The same page says that Cowork bills by usage in Copilot Credits. The Copilot Studio pricing page sells a pack of 25,000 credits for $200 a month. The billing documentation says that a spending policy always uses prepaid credits first.

    The admin documentation describes the default for automated tasks. Cowork asks the user before such a task sends an email, posts a message or changes a shared system. In the browser, Cowork shows an approval card before actions that submit personal data or that are sensitive. For passwords and verification codes, it gives the browser back to you.

    The privacy documentation says that prompts, responses and data from Microsoft Graph are not used to train foundation models. It describes classifiers against cross-prompt injection attacks. It also says that these classifiers may not be available in all Copilot scenarios.

    What I did not find: a free plan and a price for one task. The trial lasts one month and needs a card. When a user reaches the credit limit, access stops for the rest of the month. An admin must turn on usage billing before anyone can use Cowork.

    2

    ChatGPT Work

    The only agent in the top five that runs on a free plan

    Our score: 9 of 10

    Suits: one person or a team that wants finished documents, sheets and slides from a goal.

    What I verified: the ChatGPT pricing page says that ChatGPT Work is in every plan: Free, Go, Plus, Pro, Business, Edu and Enterprise. Free costs $0, Go $8 and Plus $20 a month. Pro starts at $100. Business costs $20 per user a month billed yearly, or $25 billed monthly.

    Work shares its usage limits with Codex. On Plus, the page estimates 15 to 160 GPT-6.1 Sol messages in each five-hour period. OpenAI launched Work on July 9, 2026, with the GPT 5.6 models, as Android Authority reported.

    The approvals and security page says that destructive tool calls always need approval, unless the tool has a read-only label. It names the risk directly: "Prompt injection can cause the agent to fetch and follow untrusted instructions." Web search uses cached results by default, and a safety monitor can pause a task.

    The cloud security page says that OpenAI does not use business data to train its models by default. For the full map of OpenAI agent products, read what OpenAI agent products do and cost.

    What I did not find: the training setting for Free and Plus users. The OpenAI help center blocked my reader, so I gave 1 point and not 2. Also, the message range on Plus is wide, so a long task can use your allowance quickly.

    3

    Claude Cowork

    Clear permission modes, and a plain warning about the risk

    Our score: 8 of 10

    Suits: people who work with long documents, files and spreadsheets on a desktop computer.

    What I verified: the Claude pricing page lists Pro at $20 a month, or $17 a month billed yearly. Max costs $100 and Ultra costs $200 a month. A Team Standard seat costs $25 a month, or $20 billed yearly.

    The Cowork start guide says that Cowork is for paid plans only. The desktop app runs on macOS and Windows, with Linux in beta. Cowork has three modes. Manual mode stops for approval on each action. Auto mode checks each action and blocks unsafe ones, and Skip mode runs with no checks.

    In every mode, Cowork asks your permission before it deletes a file permanently. The safety article says that web content is a primary path for prompt injection attacks. It also gives a plain warning: "the chances of an attack are still non-zero."

    Anthropic lets Free, Pro and Max users choose if their chats train its models, in its consumer terms. If you allow it, Anthropic keeps the data for five years. If you do not, it keeps it for 30 days. Team and Enterprise plans are outside that choice.

    What I did not find: a free plan with Cowork. Also, the safety article says that in Skip mode nothing checks the actions. Do not let a team use Skip mode on a task that touches accounts or money.

    4

    Gemini Spark

    A background agent for Google apps, with a second model that checks each action

    Our score: 8 of 10

    Suits: people who work in Gmail, Calendar, Docs and Drive and want tasks to run on a schedule.

    What I verified: Spark started with Google AI Ultra in the US. In July 2026 it reached AI Pro subscribers in the US, as Android Authority reported. That report says that Spark is not available in the European Economic Area, the UK, Switzerland or Nigeria.

    On July 30, 2026, Google announced AI Pro access in over 160 more countries and a Chrome connection, first in the US. Spark can use your signed-in accounts and saved passwords only with your permission. Google says that it hands sensitive actions, such as payments, back to you.

    The same post says that the Chrome connection protects against prompt injection, and it links to a Chrome security post. That post calls indirect prompt injection the primary new threat to all agentic browsers. One defense is a second Gemini model that checks each action and sees only data about the action, not the web page.

    Chrome Unboxed reported AI Pro at $20 a month in the US. The Gemini plans page showed my reader a European version, with AI Pro at 21.99 euros a month. The Gemini privacy hub says that Google uses your activity to train generative AI models, unless you turn off Keep Activity.

    What I did not find: a free plan with Spark, and a US price on a Google page that loaded for me. 9to5Google also reported a gap. Spark asks you to confirm edits to shared documents, but it can make bulk changes to your private Google Tasks without confirmation.

    5

    OpenClaw

    Free and open source, with the approval switch off by default

    Our score: 8 of 10

    Suits: a technical person who wants to run an agent on their own hardware and select the model.

    What I verified: the GitHub repository shows the MIT license and about 394,500 stars when I read it. OpenClaw runs on your own hardware, and its memory, state and credentials stay local. You talk to it through chat apps such as WhatsApp, Slack, Teams and Telegram.

    The prompt injection page is one of the clearest in this list. It says: "The content itself is a threat surface, not just the sender." It recommends a reader agent with no tools for untrusted text. It also tells you not to use weaker or smaller models for agents with tools.

    The cost: the exec approvals page says that the ask mode is off by default. With that default, ordinary commands run with no prompt to you. The page also says that after you approve a command, it can change files within the permissions of the host.

    What I did not find: a training rule, because OpenClaw does not hold the model. The training rule is the rule of the model provider that you connect, so I gave 1 point. The software is free, and you pay that provider for each token.

    6

    Meta Muse

    A free personal agent with a guard on each outside action

    Our score: 7 of 10

    Suits: people in the US who want help with errands: email, travel, forms and shopping.

    What I verified: Meta launched Muse on September 8, 2026, in the US, on iOS, Android, the web and WhatsApp. Muse runs on a dedicated cloud computer for each user, which Meta calls the Muse Secure VM. A separate agent, the Sentinel, controls what leaves that computer.

    Nothing Muse does reaches the internet unless the Sentinel approves it.

    Meta, Introducing Muse

    Meta says that Muse checks with you before it sends an email or makes a purchase. Muse cannot see your passwords or payment methods. Checkout uses Link by Stripe, which makes a card number for one use only. You get a record of past and planned actions, and you can opt out of model training.

    What I did not find: prices. Meta says that Muse is free for most of what people need, with paid plans for more, but it gives no figures. The post also does not name prompt injection. The Sentinel is a control on outside actions, so I gave 1 point for that check.

    7

    Perplexity Computer

    A cloud worker on credits, from a vendor that calls injection unsolved

    Our score: 7 of 10

    Suits: research work with many sources, long reports and scheduled checks.

    What I verified: the Computer help page describes an agent in an isolated cloud sandbox for paid plans. Some free accounts get a one-time allowance of Computer turns. The Max page lists $200 a month or $2,000 a year.

    The credits page says: "Today, 100 credits equals $1." Max includes 10,000 credits each month, and Pro has a weekly allowance with no monthly credits. When credits run out, a task pauses.

    For its browser, Perplexity lists confirmations before it sends an email or a message, changes your calendar or places a final order. The same post says that malicious prompt injection remains "an unsolved problem across the industry". The data page says that training use is on by default for Free, Pro and Max, and that you can opt out.

    What I did not find: a general approval step for Computer itself, so I gave 1 point. The help page says that Computer does not undo changes that it already made in connected apps. I also did not find the Pro price on a Perplexity page that I read.

    8

    Manus

    A large free allowance, and little public detail on controls

    Our score: 5 of 10

    Suits: people who want one prompt to become a report, a website or slides, and who will check the result.

    What I verified: the pricing article gives free users 300 credits each day. Pro starts at $20 a month for 4,000 credits, or $40 for 8,000 credits. Team starts at $20 per seat, and yearly billing gives a 17% discount.

    The cloud browser documentation asks you to take control for text message codes, CAPTCHA tests and two-factor sign-ins. It says that you control which accounts Manus can reach. For banking sites, it tells you to use your own browser, not the cloud browser.

    What I did not find: a confirmation before purchases or sent messages, a prompt injection defense and a training rule. The help article about model training returned a 404 error. Reports about the ownership of Manus disagree, so I do not state it.

    9

    Opera Neon

    No training on your data, and no public approval rules

    Our score: 4 of 10

    Suits: people who want an agent inside the browser and a clear rule about their data.

    What I verified: the Opera Neon page lists $19.90 a month, and the subscription turns on the agent functions. It says that your passwords, card details and cookies stay on your device. It also says that Neon does not use your chats or browsing history for model training.

    What I did not find: a free plan for the agent, an approval step and a defense against prompt injection. The page says that Neon can take direct action for you, but not when it stops to ask.

    10

    Alexa+

    Free with Prime, and no public detail on agent controls

    Our score: 3 of 10

    Suits: households in the US that have Alexa devices and a Prime membership.

    What I verified: Amazon says that Alexa+ is free with Prime and costs $19.99 a month without Prime. A text chat is free for everyone, with limits based on use. Alexa+ can order takeout, make reservations, book rides and schedule home repairs.

    What I did not find: a statement about confirmation before orders, prompt injection or training on the Amazon page that I read. Other Amazon help pages did not load for me. A low score here means low public evidence about the agent controls. It does not mean a bad product.

    What the scores show

    Four patterns came out of the ten sets of pages. Each one changes how I tell a client to select an agent.

    First, the large vendors agree on when to stop. Microsoft, OpenAI, Anthropic, Google and Meta all document an approval step that is on by default. The same actions appear again and again in their lists: a purchase, a payment, a sent message, a deletion and a sign-in.

    That agreement is useful to you. If an agent does not stop for those five actions, its controls are weaker than those of the large vendors, whatever its demo shows.

    Second, the training default is where they differ most. Only Microsoft and Opera say that they do not train models on your data by default. Anthropic, Google, Perplexity and Meta give you a setting to opt out. Perplexity adds that an opt-out applies only to data after the date of the opt-out.

    Third, a free plan and strong controls rarely come together. ChatGPT Work is the only agent in the top five with a free plan. OpenClaw is free too, but its approval switch is off by default. Muse and Manus are free for most use, and both publish less about their controls.

    Fourth, the scores are close at the top. Ranks 1 to 5 have 8 or 9 points. Do not read that order as a quality ranking. Read it as a list of what each vendor lets you check before you give the agent your accounts.

    Therefore: inside the top five, select by the apps that you already use. Microsoft 365 points to Copilot Cowork, Google Workspace points to Gemini Spark, and a person with no suite can start free with ChatGPT Work.

    What one person and one office pay

    The entry price tells you how to start. The meter tells you what heavy use costs. These agents use four kinds of meters: plan limits, credits, tokens and nothing that the vendor states.

    AgentLowest price that includes the agentWhat meters your use
    ChatGPT Work$0 on Free, or $20 a month on PlusUsage limits shared with Codex
    Claude Cowork$20 a month on Pro, or $17 a month billed yearlyUsage limits of the plan
    Gemini SparkGoogle AI Pro, reported at $20 a month in the USCompute-based usage limits, as in the rest of Gemini
    Meta MuseFree for most usePaid plans for more use, with no published price
    ManusFree with 300 credits a day, or Pro from $20 a monthCredits
    Opera Neon$19.90 a monthNot stated on the page
    Alexa+Free with Prime, or $19.99 a monthUnlimited, per Amazon, with limits on the free chat
    OpenClaw$0 for the softwareTokens, paid to the model provider
    Perplexity ComputerMax at $200 a month (Pro has a weekly allowance)Credits, at 100 credits for $1
    Microsoft Copilot CoworkCopilot Business at $21 a user a month billed yearly, plus a Microsoft 365 Business planCopilot Credits, at $200 for a pack of 25,000

    For one person, the arithmetic is short. Eight of the ten have a way to start at about $20 a month or less. Perplexity Max costs ten times that, and Copilot Cowork needs a business plan first.

    Look at the Perplexity row once more. Max costs $200 and includes 10,000 credits each month. At 100 credits for $1, those credits have a value of $100, which is half of the price.

    Now a scenario for an office. Consider a team of 10 people that wants one general agent for research, documents and email drafts. I assumed these inputs, so replace them with your own.

    Option for 10 peopleArithmeticMonthly cost
    ChatGPT Business, billed yearly10 seats at $20$200
    Claude Team Standard, billed yearly10 seats at $20$200
    Microsoft Copilot Business add-on, billed yearly10 seats at $21, plus one pack of 25,000 credits at $200$410, plus the Microsoft 365 Business plan
    A custom agent from Frenchy Digital$40,000 in year one, divided by 12About $3,333 in year one, plus usage

    Do the Microsoft row by hand. Ten seats at $21 cost $210. Microsoft tells admins to set a budget of at least 2,000 credits for each user each month. If each person uses that amount, the team uses 20,000 credits, which fit in one $200 pack. The total is $410, before the Microsoft 365 Business plan.

    At the pack price, one credit costs $200 divided by 25,000, which is $0.008. So 2,000 credits cost $16 for each person. If the team uses fewer credits, the pack still costs $200. With the promotional seat price of $18, the total falls to $380.

    Compare the extremes. ChatGPT Business costs $200 a month, which is $2,400 a year. A custom agent from us costs $40,000 in year one. That is about 16.7 times more.

    For this scenario, I tell the owner to use an agent from the list. A custom agent is worth its price only when the agent must write into a system that none of these agents can reach.

    Therefore: before you compare plans, count two numbers. Count the people who need the agent, and the tasks that each person gives it in a week. Then price one month in the meter of each vendor.

    What an agent may do alone

    Before you connect an agent to your accounts, write down three lists. List what it may do alone, what needs your approval and what it must never touch. It takes about an hour, and the approval step in the product is only as good as this list.

    AreaMay act aloneNeeds your approvalMust never touch
    ResearchSearch the web, read public pages, write a summary with sourcesSign in to a paid service with your accountFiles from unknown sites that it must open or run
    Documents and filesWrite drafts of documents, sheets and slides in one work folderEdits to shared documents, moves of filesPermanent deletion, files outside the work folder
    Email and messagesRead, sort and write draftsEach message that it sends, each reply to a new contactMessages to a large list, messages about money or contracts
    CalendarSuggest times and hold a slot for youAccept, move or cancel a meeting with other peopleChanges to the calendars of other people
    PurchasesCompare prices, fill a cartThe final order, a booking with a chargeSaved card numbers, bank transfers, new accounts
    AccountsRead the status of an accountA sign-in with a saved passwordChanges to passwords, permissions or recovery details

    The vendors agree with the middle column. The Chrome security post names three cases. The agent must ask before it goes to banking or medical sites, uses a saved password or completes a purchase. Meta names a sent email and a purchase. Microsoft names a sent email, a posted message and a change to a shared system.

    Perplexity names messages, calendar changes and final orders. The four lists differ in detail, but each one names sent messages, and three name purchases. Use what they share as your starting point.

    The middle column is where most of the value is. The agent does the slow part, and you approve the last step. I explain how to design that approval so that it does not become automatic in human-in-the-loop design for AI agents.

    Actions can move from the middle column to the left column, one at a time. An action that you approved 50 times with no change is a candidate. An action that you corrected one time in ten is not.

    Therefore: complete the three columns before you select an agent. Then check that the agent stops for each action in the middle column, on the default setting.

    What fails first

    Five failures occur often enough that you must plan for them. For each one, the table shows how you see it and what limits the damage.

    The failureHow you see itWhat limits the damage
    The agent follows instructions hidden in a page or an emailAn action that you did not ask for, after the agent reads outside contentApproval steps on, a separate browser profile, access only to the accounts that the task needs
    The agent changes many items at one timeMany edits in a shared document or a list, in one runReview the request before you send it, and test on a copy first
    The agent cannot reverse a changeA wrong record stays in a connected app after the task endsAsk for a plan first, split the work into small tasks
    The credits or the budget run outThe task pauses, or the agent stops for the rest of the monthA monthly limit, an alert before the limit, a manual plan for urgent work
    Someone turns the approvals off for speedThe agent finishes tasks with no prompt to youA written rule about which modes the team may use, and a check each month

    The first row has a public example. In August 2025, Brave published research on the Comet browser from Perplexity. Researchers hid instructions in a Reddit comment. When a user asked Comet to summarize the page, the agent followed the hidden text.

    The hidden text told the agent to get the email address of the user and a one-time code from Gmail. Then it told the agent to post both in a reply. With those two items, an attacker could take over the account. Brave reported the flaw on July 25, 2025, and retests found that the first fix was incomplete.

    I do not conclude that Comet was careless. Perplexity published its defenses later, and it says that the problem is unsolved across the industry. I conclude that every agent that reads the web can read an attack.

    OWASP puts prompt injection first on its Top 10 for LLM applications. The same list includes excessive agency: an agent with more permissions than its task needs. A general agent with access to your inbox and your saved passwords is the clearest case of that risk.

    The second and third rows come from the vendors. 9to5Google reported that Spark can make bulk changes to private Google Tasks without confirmation. Perplexity says that Computer does not undo changes that it already made in connected apps. I list the controls that I recommend first in ten security controls before you connect your data.

    The fourth row is specific to meters. Perplexity pauses a task when credits run out. Microsoft stops access to Cowork for the rest of the month when a user reaches the limit. Plan a manual way to finish urgent work.

    The fifth row is easy to miss, because nothing breaks. Claude has a Skip mode, and OpenClaw ships with its ask mode off. A team that wants speed can turn the approvals off and forget it. Check the mode of each agent once a month.

    The disclosed pick: Frenchy Digital

    This section is an advertisement, and I label it that way. Frenchy Digital is my company. I did not score it in the list above, because a company cannot score itself fairly.

    Our disclosed pick: Frenchy Digital

    These are the facts that you can check. We started to operate in 2016. An agent from us costs $5,000 plus a $5,000 setup fee, so $10,000 to start. After launch, it costs $2,500 a month.

    You pay model usage directly to the provider, so we add no markup to usage. Year one costs $40,000 plus usage. Each later year costs $30,000 plus usage.

    We sign business associate agreements with healthcare clients. We hold no SOC 2, ISO 27001 or HITRUST attestation. If your procurement needs one of those, select a vendor that has one.

    Our public agent evidence is the Beyond Points case study. Its chat can call 13 tools, from award search to transfer math. Transfers, bookings, payments and cancellations never run without a checkpoint from the user. That is the valet key, built into the product.

    The same case study reports an audit of the claims in the product. It found 112 timing claims, and 83 of them had no label. We examined each one, so that the interface says only what the logs support. I prefer that you read that audit, not a result number that I selected.

    Where we fit: the agent must work inside your own product, or write into a system that none of the ten agents can reach. Where we do not fit: a task that one of the ten can do on its free plan or its $20 plan. The arithmetic in the price section shows why.

    If your company is in New York, our team for AI agent development in New York can review your task with you. Anywhere else, book a discovery call. If an agent from the list fits better, I will say so.

    Warning signs when you select

    You can buy an agent or hire a builder, and that includes us. In both cases, these signs tell me not to buy.

    • No page that lists the actions where the agent stops for your approval. Ask for that list in writing.
    • An approval step that you must turn on yourself. The safe setting must be the default setting.
    • No mention of prompt injection anywhere in the documentation. A vendor that does not name the risk probably did not design for it.
    • A training rule that you cannot find in two minutes. Ask whether your chats train a model, and how you turn it off.
    • A credit meter with no example of the cost of one task. Ask for the credits that a typical task of yours uses.
    • An accuracy or success percentage with no method. Ask for the sample, the period and the definition of success.
    • A request for all of your accounts on the first day. The first task needs only the accounts that it uses.
    • Security badges that the vendor cannot prove. Ask for the report, not the logo.

    On the last point, this is our own position, because I ask the same of other vendors. If your procurement process needs an attestation, a vendor that has one will pass, and we will not. You must know that before the first call.

    Limitations

    This is what I did not verify, or decided not to use as evidence.

    • I read each page on November 15, 2026. Vendors change prices, plans and features often, and several of these agents launched in 2026.
    • Some pages blocked my reader: the OpenAI help center, the Genspark pricing page and some Amazon help pages. I left Genspark out, and I gave OpenAI 1 point on training for that reason.
    • The Gemini plans page served my reader a European version in euros. The US price of $20 for AI Pro comes from a news report, not from a Google page.
    • The Spark availability by country comes from a news report. The Google post confirms only the 160 added countries for AI Pro.
    • A zero means that I did not find the fact on the pages that I read. It does not mean that the feature does not exist.
    • The scores measure public evidence, not quality. An agent with 5 points can be the correct choice for your task.
    • I did not test the agents with live accounts for this article. The scores come from documentation, not from a trial.
    • The office scenario uses assumed inputs. The credits that your own team uses depend on its tasks.
    • Our own public agent evidence is one case study. That is a small sample.

    Three tasks for this week

    1. 1.Select one task that you do every week and that has a correct result. Write down five past examples with the correct result for each one.
    2. 2.Complete the three columns for that task: may act alone, needs approval, must never touch.
    3. 3.Start the free plan or the $20 plan of the top agent that fits your apps. Run your five examples with the approval step on, and count the results that you can use.

    If the task must write into a system that none of these agents can reach, book a discovery call with us. If an agent from the list fits, I will tell you.

    Start with five examples.

    Which Task Do You Give an Agent First?

    Book a discovery call with Frenchy Digital, a senior-led, Black-owned agency in Los Angeles. Bring one task, the apps that it uses and the data that it touches. If an agent from this list fits, we will tell you. If you need a custom agent, you get a fixed-price proposal.

    Do You Want a Second Opinion on Your First Agent?

    Book a discovery call. Bring one task, the apps that it uses and the data that it touches. We will tell you which agent fits, or if you need a custom one.

    1517 S Bentley Ave Suite 204, Los Angeles CA 90025

    Frequently Asked Questions

    Sources & References

    1. 1Microsoft Learn: manage Copilot Cowork for your organization↗
    2. 2Microsoft Learn: use the local browser with Copilot Cowork↗
    3. 3Microsoft Learn: managing AI experiences enabled by usage-based billing↗
    4. 4Microsoft Learn: data, privacy and security for Microsoft Copilot↗
    5. 5Microsoft: Microsoft 365 Copilot pricing for business↗
    6. 6Microsoft: Copilot Studio pricing (Copilot Credit packs)↗
    7. 7Redmond Channel Partner: Microsoft announces general availability of Copilot Cowork (June 16, 2026)↗
    8. 8OpenAI: ChatGPT plans and pricing for Work and Codex↗
    9. 9OpenAI: get started with ChatGPT Work↗
    10. 10OpenAI: agent approvals and security↗
    11. 11OpenAI: ChatGPT Work cloud security↗
    12. 12Android Authority: OpenAI launches ChatGPT Work with GPT 5.6 (July 9, 2026)↗
    13. 13Anthropic: Claude plans and pricing↗
    14. 14Anthropic Help Center: get started with Claude Cowork↗
    15. 15Anthropic Help Center: use Claude Cowork safely↗
    16. 16Anthropic: updates to our consumer terms and privacy policy (August 28, 2025)↗
    17. 17Google: Gemini Spark now integrates with Chrome (July 30, 2026)↗
    18. 18Google: architecting security for agentic capabilities in Chrome (December 8, 2025)↗
    19. 19Android Authority: Gemini Spark wider rollout (July 24, 2026)↗
    20. 209to5Google: Gemini Spark rolls out to Google AI Pro in the US (July 23, 2026)↗
    21. 21Chrome Unboxed: Gemini Spark is now available on the $20 Pro tier in the US (July 27, 2026)↗
    22. 22Google: Gemini plans↗
    23. 23Google: Gemini Apps Privacy Hub↗
    24. 24OpenClaw: GitHub repository↗
    25. 25OpenClaw documentation: exec approvals↗
    26. 26OpenClaw documentation: prompt injection↗
    27. 27Meta: introducing Muse (September 8, 2026, updated November 10, 2026)↗
    28. 28Perplexity Help Center: what is Computer?↗
    29. 29Perplexity Help Center: how credits work on Perplexity↗
    30. 30Perplexity Help Center: Perplexity Max↗
    31. 31Perplexity: mitigating prompt injection in Comet (October 22, 2025)↗
    32. 32Perplexity Help Center: data collection at Perplexity↗
    33. 33Brave: agentic browser security, indirect prompt injection in Perplexity Comet (August 20, 2025)↗
    34. 34Manus Help Center: current membership pricing↗
    35. 35Manus documentation: cloud browser↗
    36. 36Opera: Opera Neon↗
    37. 37Amazon: Alexa+ now available to everyone in the US, and free for Prime members↗
    38. 38OWASP GenAI Security Project: Top 10 for LLM applications↗
    Chris Machetto - CEO & Founder, Frenchy Digital of Frenchy Digital

    Chris Machetto

    CEO & Founder of Frenchy Digital. Building apps and digital products since 2016 for startups and enterprises across LA, San Francisco, Paris, Geneva, and more globally.