The agents are here. Here's what they actually do.

Launched on 9 July 2026 as one of three modes, alongside Chat and Codex, in a new ChatGPT desktop app, and built on Codex. It breaks a project into steps, works through connected apps such as Slack, Google Drive and calendars, a built-in browser and, on desktop, computer use, and produces documents, spreadsheets, slides and small web apps. It can run once, on a schedule or when something changes, and you can watch, redirect it and approve important actions.

TechCrunch's reviewer found setting up permissions "confusing and circular", could not grant read-only access to a cloud drive, and estimated that four days of casual use on a $20 plan burned about $65 of compute. No Remote Labor Index score has been published for Work; its predecessor, ChatGPT agent, automated 1.25% of projects.

Cowork launched on 12 January 2026 as a research preview for Max subscribers on macOS, built on the Claude Code harness for non-coding work. You choose which folders and tools it can reach; it plans, works through the steps in view, can use sub-agents and a built-in browser, and can run scheduled tasks unattended. Since 6 October 2026 new tasks on Pro, Max and Ultra run in the cloud, reaching local folders only through the desktop app while it is open.

The defaults are careful ("Deleting anything needs your approval"), but computer use is still a research preview and Anthropic warns that Cowork takes real actions on your files. On the Remote Labor Index, Claude Opus 4.6 in the Cowork harness automated 4.17% of projects.

Gemini Agent arrived in the Gemini app on 18 November 2025 for US AI Ultra subscribers, drawing on Project Mariner to manage calendars, reminders and inboxes. At I/O on 19 May 2026 Google announced Spark, "a 24/7 personal AI agent that helps you navigate your digital life", which runs in the cloud so it keeps working when your laptop is closed. In October 2026 Google began folding Spark into a "Task" mode beside chat, and de-emphasising the Spark name.

It is designed to ask first before high-stakes actions such as spending money or sending email. We found no independent test of how much it finishes on its own; Google's strength is that it already holds your email, calendar and documents.

Genspark's Super Agent arrived in April 2025, with a demo in which a synthetic-voice agent phoned restaurants to book a five-day trip. Its AI Workspace, launched in November 2025, takes a high-level goal and picks among more than 30 models and 150 in-house tools to deliver documents, financial models, presentations and apps without step-by-step prompting. The company was valued at $2.6 billion in June 2026.

Its co-founder Eric Jing said at launch that "solving these real-world problems is much harder than we thought". We found no independent evaluation of its success rate; its revenue figures are self-reported.

Each Bot works on a cloud computer, so it can sign in and operate apps, inboxes and websites that have no clean API. You message it like a colleague; it remembers conversations, can save a workflow you demonstrate as a routine, and several Bots can coordinate in a group chat. The beta launched on desktop and iOS on 11 August 2026, widened to more plans on 26 August, and an enterprise version with access, network and audit controls followed on 3 September.

It is an early beta. xAI says Bots pause for approval when needed, but its launch post gave few details on security, and we found no independent test of how reliably they finish work.

Manus launched as an invite-only beta on 6 March 2025, letting you watch the agent work in a "Manus's Computer" window and step in. Meta agreed to buy it in December 2025, Chinese regulators blocked the deal in April 2026, and Manus announced on 11 August 2026 that it was independent again. Manus 2.0, released on 28 September 2026, added a new harness called Cascade, automations triggered by events in connected services, dedicated cloud computers and control of your own computer from your phone.

On the Remote Labor Index, Manus 1.6 Max automated 2.92% of projects, below the frontier labs' best. Manus's claims that 2.0 uses fewer tokens and costs less come from its own testing.

Computer breaks an outcome into tasks and spins up sub-agents for web research, documents, data processing and calls to connected services. Each task runs in an isolated environment with a file system and browser. At launch Claude Opus 4.6 coordinated sub-agents using Gemini, Nano Banana, Veo 3.1, Grok and GPT-5.2, and it can run for long periods and in parallel.

Perplexity says it will "check in" when it truly needs you, but describes few other safeguards. We found no independent evaluation of how well it finishes work.

Copilot Tasks takes a request in plain language and works through it in the background on its own computer and browser: drafting email replies, planning events, making documents, tracking listings or booking services. Tasks can recur on a schedule. It asks for confirmation before spending money or sending messages, and you can review or cancel tasks.

It launched as a research preview for a limited group, with a waitlist. We found no reporting on wider availability or pricing since.

Muse works through connectors to Gmail, Google Calendar, Instagram and, in Singapore, partners such as Shopee, Lazada, Grab, Singapore Airlines, Agoda, Klook, Trip.com and ShopBack. Meta says it checks with you before sensitive actions such as sending an email or making a purchase, and that it runs on a dedicated cloud computer (Muse Secure VM) that keeps passwords and payment methods out of the agent's view.

Meta says connector features will roll out progressively, with some experiences arriving after launch, and AI-glasses support has no date. We found no independent test of how reliably it completes bookings or purchases.

The r1 launched in April 2024 promising a Large Action Model to operate apps. In October 2024 Rabbit opened a beta LAM Playground, a 'generic website agent' using a vision model to work through pages. Rabbit also sells rabbit intern, a separate all-purpose agent that produces websites, reports and documents.

A review updated in August 2026 describes the agent as still not the Large Action Model the launch promised, with narrow app coverage and slow responses. Rabbit's own beta notes cited bugs and hallucination. We could not confirm from Rabbit's pages how intern relates to the r1.

Released in beta with iOS 27 on 14 September 2026, on models Apple built with Google's Gemini, Siri AI can draft email, edit and share photos, add events and reminders from what is on screen, and build Shortcuts from a description. It works with some third-party apps, including WhatsApp and Audible. Access came through a waitlist, and it needs an Apple Intelligence device such as an iPhone 15 Pro or later.

It acts one request at a time inside apps that support it, and few third-party apps did at launch. Apple says the Digital Markets Act is why there is no timeline for it on iPhone and iPad in the EU.

Lindy reads your connected inboxes, drafts replies, schedules and prepares meetings, takes notes and handles follow-ups, through iMessage, email, Slack or the browser. Its pricing page says 'Anything with outside impact waits for approval', so sends and publishes are approved by you; computer use is offered on higher plans.

Usage is metered in credits, which do not roll over, and it needs broad access to your email, calendar and documents. Its own launch page says you review drafts before anything goes out. We could not find an independent evaluation of its error rate.

Atlas put ChatGPT beside every page and offered an agent mode that could open tabs, click through sites and complete tasks such as hotel bookings, asking for confirmation before sending messages or making payments. OpenAI announced its retirement on 9 July 2026, alongside ChatGPT Work, and scheduled the shutdown for 9 August 2026.

It never left the Mac. OpenAI said its agent mode "expands the security threat surface" and wrote that prompt injection "is unlikely to ever be fully 'solved'".

Launched in preview on 28 January 2026, auto browse compares hotel and flight prices, fills forms (including from a PDF) and adds items to a basket with discount codes. With permission it can use Google Password Manager to sign in. It is designed to "pause and explicitly ask for your confirmation" before sensitive steps such as a purchase or a social post.

It is a US-only preview behind a paid plan. We found no independent test showing how often it finishes a task unaided.

Piloted with 1,000 Max subscribers in August 2025, it became generally available on every paid plan on 26 August 2026. Since 12 August 2026 its side panel runs as a Cowork session, so skills and connectors work in the browser and a task started in a tab can be finished in the desktop, web or mobile apps. With auto-approve on, a classifier checks consequential actions such as submitting forms or sending messages and blocks those that don't match the request; it still asks before purchases or sharing personal data.

It works with your logged-in sessions, so prompt injection is the main risk. Anthropic's own tests found attacks succeeded 23.6% of the time in the 2025 pilot (11.2% with mitigations); in 2026 it reports none succeeding against Sonnet 5, Opus 5 and Mythos 5 with safeguards on, and 0.3% against Fable 5. These are vendor tests, and the method changed in between.

Comet launched on 9 July 2025 for top-tier subscribers and became free to download on 2 October 2025, followed by Android in November 2025 and iOS in March 2026. Its assistant can act on the page you are on, background assistants can work on several tasks at once, and an Email Assistant for Max subscribers schedules meetings and drafts replies.

It became security researchers' favourite target. Brave called indirect prompt injection "a systemic challenge facing the entire category of AI-powered browsers", and showed hidden text on a web page could take over Comet's assistant.

Unveiled in May 2025, Neon shipped to early users in autumn 2025 and opened to the public on 11 December 2025 as a $19.90-a-month subscription. Its agents can carry out tasks on pages, run deep research and build mini apps, and "Tasks" group chats and tabs into workspaces. Since August 2026 the browser itself is free, and outside agents can drive it through MCP and a command-line tool; the paid plan unlocks Neon's own autonomous browsing.

We found no independent test of how reliably its agents complete tasks.

Deep research browses for roughly five to thirty minutes, reading and cross-checking sources, and writes a cited report. A February 2026 update moved it to a GPT-5.2-based model, let users restrict it to chosen sites, connect apps and data through MCP servers, and steer it mid-run. When OpenAI published its BrowseComp benchmark in April 2025, deep research solved 51.5% of its hard-to-find questions; the paper noted it had been trained for that kind of task.

OpenAI acknowledges it sometimes makes factual errors and wrong inferences. An April 2026 study found that 3–13% of citation URLs from commercial models and deep research agents had no record in the Wayback Machine and likely never existed, and that deep research agents fabricated URLs at higher rates.

Deep Research launched for Gemini Advanced subscribers on 11 December 2024. On 11 December 2025 Google released a stronger version built on Gemini 3 Pro, available to developers through the Interactions API and as an upgrade in the Gemini app, and claimed 59.2% on OpenAI's BrowseComp benchmark.

The benchmark figure is Google's own. An April 2026 study found that 3–13% of citation URLs from commercial models and deep research agents likely never existed, with deep research agents fabricating at higher rates.

Launched on 14 February 2025, Deep Research runs dozens of searches, reads hundreds of sources, refines its plan as it goes and writes a report that can be exported or turned into a Perplexity Page. Perplexity claimed 21.1% on Humanity's Last Exam at launch. In February 2026 it moved to Claude Opus 4.6, first for Max users.

Its benchmark figures are Perplexity's own. An April 2026 study found that 3–13% of citation URLs from commercial models and deep research agents likely never existed, with deep research agents fabricating at higher rates.

Announced on 9 March 2026 and built "in close collaboration with Anthropic" on the same harness as Claude Cowork, it takes a goal such as preparing a meeting pack, plans the steps and carries them out across Microsoft 365, running in the cloud so tasks continue when your laptop is off. It became generally available on 16 June 2026, running on Anthropic models, and absorbed the computer-use feature from Microsoft's Researcher agent, which was retired on 1 July.

Each task's price depends on model use, tool calls and run time, so costs are hard to predict; Microsoft's claim that it is 30–40% cheaper than Claude Cowork comes from its own testing. Web browsing was limited to the Frontier programme at launch.

Notion 3.0, released on 18 September 2025, rebuilt Notion AI as agents. Notion says anything you can do in Notion, your agent can do, working for over 20 minutes across many pages with a memory kept in Notion pages. Custom Agents that run on schedules or triggers were announced as coming soon.

Notion's release page does not say which plans include agents, and gives no usage quotas. Custom Agents were not yet available at launch; later release notes suggest they have since shipped, but we have not verified their current terms.

Reclaim works inside Google Calendar or Outlook Calendar. It places tasks, habits and focus time into free slots by priority, finds meeting times across attendees, and adapts the plan when your schedule changes.

It manages time, not email or other apps, and it works only on the calendars you connect. Dropbox did not disclose the terms of its acquisition.

Sidekick works inside the Shopify admin with the permissions of the logged-in user. In December 2025 Shopify added the ability to build custom admin apps and visual automation workflows from a prompt, and to save reusable prompts as 'skills'. Shopify's developer documentation says Sidekick suggests changes and the merchant confirms them.

Shopify's docs say Sidekick does not change the store on its own from a launched prompt, and it is not available on every store. A retail analyst quoted by Retail Brew said merchants will want reliability to improve before trusting it beyond low-stakes tasks. Launch date and pricing not verified.

Claude Code searches and reads a codebase, edits files, writes and runs tests, uses command-line tools and can commit and push to GitHub. It began as a limited research preview in February 2025 and became generally available that May. Anthropic's newest model, Claude Opus 5.5 (22 September 2026), comes with a classifier that Anthropic says screens every action before it runs.

It acts with whatever permissions it is given, so a mistaken command can be destructive, and its own safety claims come from Anthropic's tests rather than independent ones. Long tasks still need a person to check the result.

Codex launched in May 2025 as a research preview of a cloud agent: each task runs in an isolated container preloaded with your repository, and the agent returns its changes with terminal logs and test output for you to review. There is also an open-source command-line version. Codex now sits in the ChatGPT desktop app alongside ChatGPT Work, and OpenAI said in June 2026 that it had more than 5 million weekly active users.

At launch it could not fully test applications with live user interfaces, and a person still has to verify and merge its work. Its usage figures are OpenAI's own and have not been independently checked.

Cursor is a code editor built around an AI agent that can change many files and run commands for you, using its own Composer models or models from other labs. SpaceX completed its all-stock acquisition of Anysphere, valued at about $60 billion, on 14 August 2026, and the team joined SpaceXAI. Cursor says the deal gives it access to SpaceX's GPU fleet to train cheaper, more capable models.

Its agent can take destructive actions with whatever credentials it can reach, and Cursor's own guidance relies on people approving privileged operations. It has also been slow to disclose what its models are built on.

Cognition introduced Devin in March 2024 as "the first AI software engineer", an agent that takes a task and works through it in its own cloud environment. Devin 2.0, in April 2025, cut the entry price from about $500 a month to $20. After buying Windsurf in July 2025, Cognition renamed that editor Devin Desktop in June 2026, so local and cloud agents can be managed from one place. Cognition raised more than $2 billion at a $48 billion valuation in September 2026.

The "software engineer" label sets expectations that an agent which still needs its work reviewed does not meet. Cognition has not published independent evaluations of how often Devin completes real tasks unaided.

Antigravity launched in November 2025 alongside Gemini 3 as a code editor with a separate manager for running several agents at once; they work across the editor, terminal and browser and report back with task lists, screenshots and recordings. Antigravity 2.0, at I/O in May 2026, added a desktop app that can schedule background tasks, a command-line tool that replaces Gemini CLI, and an SDK.

Agents that browse the web and run shell commands can be steered by what they read, and Google lists browser-agent data exfiltration through prompt injection as a known issue. Its usage limits are tied to Google's consumer subscription tiers.

Grok Build is a command-line agent that drafts a plan you can approve or rewrite, then edits files and runs shell commands, delegating parts of larger jobs to subagents that run at the same time. It can also run headless inside scripts and works with existing AGENTS.md files, hooks and MCP servers. After a privacy row in July 2026 its source code was published on GitHub under the Apache 2.0 licence.

It launched as an early beta and its data handling was found wanting within two months. The model and harness are less proven than the established coding agents.

You describe what you want and Lovable's agent builds a working web app, with hosting and a database, then changes it as you keep chatting; a plan mode lets you agree the approach first. Lovable says more than 60 million projects have been made since it launched in November 2024. It raised $400 million at a $13.3 billion valuation in August 2026.

Many of its users cannot check the security of what it builds, and twice researchers have found data exposed. Credits are spent per change, so heavy iteration costs more than the headline price suggests.

Replit Agent sets up the development environment, installs dependencies, writes and runs code and deploys the result from a short description. Agent 4, released in March 2026, added a design canvas, a planning mode, several tasks running in parallel and connections to tools such as BigQuery, Linear, Slack and Notion. Replit raised $400 million at a $9 billion valuation the same month.

It is aimed at people who may not read the code it writes, so mistakes can go unnoticed until they matter. Costs scale with the agent's effort, which is hard to predict in advance.

Formerly called Copilot coding agent, it takes a task, works on it in a separate environment and returns a branch or draft pull request for review. Since April 2026 it can also research a codebase and propose a plan for approval before it writes code. Through Agent HQ, announced in October 2025, GitHub also brings coding agents from Anthropic, OpenAI, Google, Cognition and xAI inside GitHub.

Nothing merges without a person, which is the point, but review becomes the bottleneck. GitHub said in April 2026 that its previous pricing was "no longer sustainable" and moved to metered credits, so the cost of a long session is harder to predict.

Gemini CLI brought Gemini into the terminal for coding, editing files, running commands and troubleshooting, with Google Search grounding and MCP support, under the Apache 2.0 licence. At launch a personal Google account got 1,000 free requests a day. At I/O on 19 May 2026 Google announced it was moving users to Antigravity CLI and that Gemini CLI would stop serving free users and Google AI Pro and Ultra subscribers on 18 June 2026.

For most individuals it has been retired, and Google said its successor would not have full feature parity at first. It reads project files as instructions, which made it open to prompt injection.

Jules clones a GitHub repository into a cloud virtual machine, drafts a plan with Gemini, makes the changes and presents a diff for you to approve before it opens a pull request. You can start it from a prompt or by labelling a GitHub issue. It went into public beta in May 2025 and came out of beta on 6 August 2025.

It works only through GitHub and waits for approval at the plan and the pull request, so it suits chores better than open-ended projects. Google publishes daily task caps rather than any measure of how often it succeeds.

Windsurf was an AI code editor whose agent, Cascade, edited code and ran commands. In 2025 OpenAI's reported $3 billion deal to buy it lapsed; Google then paid about $2.4 billion to license its technology and hire its chief executive, Varun Mohan, co-founder Douglas Chen and some researchers; and on 14 July 2025 Cognition agreed to buy the rest, including the product, brand and remaining staff. On 2 June 2026 Cognition renamed it Devin Desktop and replaced Cascade with Devin Local; legacy Cascade ran until 1 July 2026.

As a separate product it no longer exists; its users are now on Cognition's roadmap. Google's deal took the founders and some researchers but left most of the roughly 250-person team behind.

Agentforce lets companies build and deploy agents that answer customers and take actions on their Salesforce data; Agentforce Service Agent was the first to become generally available, in October 2024, priced at $2 per conversation. Salesforce has since added metered Flex Credits and per-user licences, and in September 2026 bought Fin to sit alongside it. MarTech reported in June 2026 that Agentforce had reached $1.2 billion in annual recurring revenue in Salesforce's first quarter of fiscal 2027.

It is strongest where a company's data already lives in Salesforce, and its pricing has changed several times, making costs hard to compare. Agents that read customer-submitted text can be manipulated by it.

Decagon's agents learn from a company's help content and past conversations and connect to its systems to act, for example processing a refund. Teams set the agents' workflows, test them in simulations and A/B experiments, and monitor them with always-on quality checks. It raised $250 million at a $4.5 billion valuation in January 2026.

Its performance claims come from the company and its customers; we found no independent evaluation. Like any deployed support agent, it speaks for the company, which is answerable for its mistakes.

Intercom launched Fin in March 2023 on GPT-4; it now runs on the company's own support model, Apex, and works with Intercom or other helpdesks such as Salesforce and HubSpot. Intercom renamed itself Fin in May 2026, and Salesforce completed its purchase of the company, agreed at about $3.6 billion, on 10 September 2026. Salesforce says Fin serves more than 30,000 companies and resolves 76% of queries on average.

The 76% figure is the vendor's own average; results depend on the quality of each company's help content. Billing per outcome counts some handoffs and disqualifications as well as resolutions.

Sierra builds branded agents that handle customer conversations and take actions in a company's systems, such as changing an order or processing a claim. Founded by Bret Taylor and Clay Bavor, it launched in February 2024, raised $950 million at a $15.8 billion valuation in May 2026, and said on 30 October 2026 that it had passed $400 million in annual recurring revenue and works with almost half of the Fortune 50. In 2024 it published τ-bench, a public test on which the best agent then succeeded less than half the time.

Its own τ-bench showed that consistency is the hard part: the best 2024 agent got the same retail task right eight times running only about a quarter of the time. When the deploying company's interests and the customer's differ, the agent serves the company, which remains responsible for what it says.

CoCounsel Legal launched in August 2025 with Deep Research, which writes a multi-step research plan, carries it out and returns a cited report grounded in Westlaw and Practical Law. The next generation, released in the US on 20 August 2026 and built on Anthropic's Claude Agent SDK, adds a brief-drafting tool, a drafting agent in Word, review of up to 10,000 documents at once and a check of whether cited authority supports each assertion.

Its answers are grounded in Thomson Reuters' own databases, which helps with citations but limits it to what those sources cover. We found no independent evaluation of its accuracy, and lawyers remain responsible for checking its work.

Harvey offers purpose-built and custom agents for work such as due diligence, contract review, drafting and legal research, including inside Microsoft Word. In March 2026 it said more than 100,000 lawyers at about 1,300 organisations used it and that over 25,000 custom agents ran on the platform. In September 2026 it raised $550 million at a $15.5 billion valuation and said 80% of Am Law 100 firms use it.

We found no independent evaluation of its accuracy, in a profession where accuracy is the product. Lawyers remain responsible for anything filed in their name.

You talk to Nous over WhatsApp and give it your bills and logins. It watches contracts and renewals, moves households to cheaper energy, broadband, mobile and car insurance deals, cancels unwanted subscriptions, fills in forms and books repairs, working within permissions you set. It says almost 250,000 UK homes use it, and it is registered with the FCA, which lets it handle regulated products such as insurance.

Its headline saving, £734 for a typical household, is the company's own figure. It uses fixed rules rather than its language models for money decisions, which it says reduces errors; how much it does without asking depends on the permissions each household gives it.

Rocket Money's negotiation team contacts providers such as internet, cable and phone companies for you and charges a share of the savings only if it succeeds. Rowan, launched on 25 August 2026 and built with Anthropic, watches your spending around the clock, texts you when a free trial is about to charge or a bill rises, and acts on a plain-language reply: cancelling a subscription, negotiating a bill or setting up savings transfers.

A one-word reply can trigger an action, so the main safeguard is the text exchange, backed by what Rocket Money calls human verification by its staff. Its negotiation fee can take more than half of the first year's saving.

Cleo is a budgeting and cash-advance app with a chatty AI persona. Autopilot, announced on 5 February 2026 for US users via a waitlist, analyses your finances, sets a plan towards a goal and gives daily guidance, acting 'always with your consent on a plan you've already approved'. At launch Cleo said the Actions component recommended moves, and carrying them out for you was a planned capability.

By its maker's own launch description the autonomous part was still to come, and we could not verify how widely Autopilot is now available or what it does unprompted.

DoNotPay says it offers more than 100 AI-powered tools. It writes and sends letters, cancels subscriptions, contests parking tickets and issues a virtual card that declines charges when a free trial ends. Its site says it is 'not a law firm and does not provide legal advice'.

Its own site says the tools are not guaranteed to help in every situation. The FTC found that it had never tested whether its 'robot lawyer' performed like a human lawyer.

Klarna announced the assistant on 27 February 2024, built with OpenAI. It handles refunds and returns, shows real-time balances and payment schedules and explains spending limits, in more than 35 languages. In September 2024 Klarna added a chat-based shopping assistant that compares products, prices, stock, delivery and cashback.

Klarna's launch figures (2.3 million chats in a month, two-thirds of service chats, satisfaction on a par with human agents) are the company's own, and the sources we opened do not say whether the shopping assistant can complete a purchase. Customers can still choose a human agent.

On 13 May 2026 Amazon replaced Rufus, its shopping chatbot since 2024, with Alexa for Shopping, which draws on Alexa+ and your shopping history. It answers product questions, compares items, rebuilds baskets from past orders, schedules recurring orders and runs conditional rules such as adding sunscreen to your basket if it falls to $10. Auto Buy, added to Rufus for US Prime members in November 2025, places the order itself once a target price is met; Amazon said it checks prices every 30 minutes.

It shops only where its maker sells, and some of its AI results are sponsored, so its advice and Amazon's interests overlap. Most actions stop at the basket for you to check out; Auto Buy orders can be cancelled within 24 hours, before they ship.

When Amazon doesn't sell an item, search results can show products from brands' own sites. You confirm the address, payment and fees on an Amazon checkout page, and Amazon's agent completes the purchase on the brand's website using your encrypted details. The brand handles delivery, returns and customer service. Alexa for Shopping now offers it too.

Amazon can't see or support the order after it is placed, so problems go to a shop you may never have visited. The listings are built from other sites' data, and merchants have reported wrong photos and details.

In Search you describe a job, such as pet grooming, and Google's AI calls several local businesses about price and availability, then sends you a summary. It began as a Search Labs test in January 2025 and reached all US Search users in July 2025; in November 2025 Google added calls to shops about stock. In September 2026 it began testing Call for Me, in which Gemini rings a business from your own number, navigates phone menus, waits on hold and can make a reservation or reschedule an appointment while you watch a live transcript.

Businesses can opt out of the calls. Google is keeping Call for Me small because, it says, 'real-world conversations are nuanced', and we found no published figure on how often calls go wrong.

Agentic checkout, rolled out in November 2025, lets you track an item with a size, colour and maximum price; when it falls within budget, Google can buy it on the merchant's site with Google Pay once you approve. Since January 2026 a 'Buy' button in AI Mode and the Gemini app checks out with eligible US retailers through Google's Universal Commerce Protocol. Since 27 August 2026 AI Mode can also book hotels from partners including Booking.com, Expedia, Hilton, IHG and Marriott.

The retailer or hotel stays the merchant of record and handles any problems, so Google takes the order but not the after-sales responsibility. The same interface that shops for you also shows advertisers' 'Sponsored deal' offers.

Launched 8 December 2025, the app lets users plan meals, build a cart from Instacart's retail partners and pay inside ChatGPT, using OpenAI's Agentic Commerce Protocol with Stripe. Users sign in to their Instacart account and approve the list before paying.

At launch payment was by credit card only, with Apple Pay, Google Pay and native mobile checkout still to come. Reports we opened did not list regions or retailers covered.

From 29 September 2025, logged-in US users on Free, Plus and Pro plans could tap 'Buy' to purchase from US Etsy sellers without leaving ChatGPT, with more than a million Shopify merchants promised to follow. It ran on the Agentic Commerce Protocol, built with Stripe. In March 2026 OpenAI moved away from it: ChatGPT now concentrates on product discovery, and purchases happen on retailers' own sites or in retailer apps inside ChatGPT.

CNBC reported that six months in, product selection was limited, item data was not always current and about 30 Shopify merchants were live. Walmart found conversion three times lower for products sold directly in ChatGPT than for those sent to its own site.

Kayak's AI Mode, powered by ChatGPT and Kayak's own data, lets users type questions and get results from hundreds of travel providers. TechCrunch reports that users can compare and book flights, hotels and cars from the chat; Skift does not say how bookings are completed.

The two reports differ on how booking works inside the chat, and we could not verify it from Kayak's own pages. At launch it was English-only, in the US, on the web rather than in the app.

Sparky arrived in the Walmart app in June 2025 to summarise reviews, recommend products and help plan purchases, with reordering and service booking promised. By May 2026 Walmart said shoppers could use it to reorder items they buy regularly, including in stores, and that it offered personalised replenishment and meal planning.

Its job is to sell Walmart's stock, and Walmart measures it by how much more its users spend: about 35% more per order, by the company's account. We found no published detail on which purchases it makes without a confirmation step.

DJ plays a stream of tracks chosen from your listening history and introduces them in a synthetic voice. Since May 2025 Premium listeners can steer it by voice with a mood, genre, artist or activity, and since October 2025 by typing too. In September 2025 Spotify added a 'Talk to DJ' button on Android Auto, and in May 2026 it expanded DJ to more than 75 countries and four more languages.

It controls one thing, the music queue: Spotify says requests for podcasts or audiobooks don't work, and it launched requests as a beta. Calling it an agent is generous, since it acts only inside the app.

Alexa+ runs on Echo devices, in the Alexa app and at Alexa.com. Amazon says it can order takeaway, find and book restaurants, book a ride, schedule services such as home repairs, shop Amazon and manage family calendars, working through partner services. In the UK the launch partners for bookings and orders were OpenTable, Just Eat and Treatwell.

It is only as capable as its partner list: in Fortune's July 2025 hands-on it could not book a restaurant that wasn't on OpenTable, and it said Spotify was playing when it wasn't. It is far more conversational than the old Alexa, but outside the US and Canada it is still in early access.

Gemini for Home replaces Google Assistant on Nest speakers and displays and in the Google Home app. You can control devices in plain language and ask it to create automations. On the top tier it describes camera events, answers questions about video history and sends a daily 'Home Brief'. Early access began in October 2025 and spread to 16 more countries in April 2026.

Its camera descriptions make confident mistakes: 9to5Google's reviewer found a disc golf cart described as a bicycle, and cited The Verge's report that it called a shotgun a 'garden tool'. Most features are still in early access.

Hippocratic AI sells a catalogue of non-diagnostic agents to providers, payors and pharma. They make calls such as post-discharge follow-ups, medication-adherence check-ins, screening outreach and intake, and escalate patients to nurses. The company says about 0.5% to 1% of live calls are sampled for safety review.

The company says its agents do not diagnose or prescribe and it will not build agents for hospice, mental-health disorders or children under two. Its safety figures (over 250 million interactions, zero reported incidents of harm) are company-reported and we have not verified them independently.

Health AI answers health questions and, if you connect your One Medical records, explains lab results and medications. It books appointments with One Medical providers by message, video or in person, and handles prescription renewals, which can be filled through Amazon Pharmacy. It began in the One Medical app in January 2026 and reached Amazon.com and the Amazon app in March.

Amazon says it is 'not intended for diagnosis or treatment without the support of a care provider' and hands clinical judgement to human providers. Every route it offers leads to Amazon's own clinics and pharmacy.

A Utah patient verifies their identity and answers a short AI interview about a maintenance medicine. Doctronic checks their prescription history, interactions and allergies, then sends the renewal to a pharmacy or refers the case to a doctor. About 190 drugs are eligible; controlled substances, injectables and short-course antibiotics are not. Under the state agreement, doctors review the first 250 renewals in each drug class before the AI may approve renewals on its own, with sampling after that.

In June 2026 the pilot was still in its first phase, with doctors reviewing every renewal. Utah's medical licensing board asked the state to pause it, and a Stanford health law analysis noted little published evidence of performance before launch and terms of service that disclaim liability.

Ebb is a conversational companion trained in motivational interviewing. It helps members reflect on their thoughts and emotions and then recommends meditation, mindfulness and wellbeing content. Headspace added voice in December 2025 and says it plans longer-term memory.

Headspace calls Ebb a 'sub-clinical support tool': it is not for clinical use and gives no mental health advice or diagnoses, and the company says AI is not a replacement for human-based care. Ebb watches for language about suicidal ideation, self-harm, abuse, eating disorders and substance abuse and then shows emergency and local resources. It is not a substitute for crisis support. We found no regulator clearance for Ebb as a medical device.

Wysa is a smartphone conversational agent that delivers cognitive behavioural therapy techniques for low mood and anxiety, alongside a self-help library and optional human coaches. Its clients have included health providers and government agencies, among them the NHS and Singapore's Ministry of Health.

The FDA gave Wysa Breakthrough Device Designation in May 2022 for a version aimed at adults with chronic musculoskeletal pain and depression or anxiety. That is a designation to speed development, not an approval, and we found no marketing authorisation. Wysa is not a substitute for crisis support or for a clinician.