Not long ago, the benchmark for useful AI was a program that could write code, schedule a meeting, or answer a customer support ticket. That standard is already starting to feel quaint. The next wave of artificial intelligence is not about producing text or completing a single task; it is about taking action on your behalf, often without you watching, and often in the messy, unglamorous corners of the real economy. This is the shift captured by Forbes’ inaugural Agentic 20 list, a snapshot of twenty companies and projects that are trying to make AI agents genuinely useful to a much larger market than software developers. The list is not a formula-driven ranking, and Forbes is upfront about that. There is no scoring methodology for a market that is still sorting itself out. Instead, the list draws on reporting, founder interviews, investor conversations and signs of actual use. The result is a portrait of an industry that is moving from the digital world of calendars and code to the physical one of shrimp farms, funeral homes, tax audits, loan disputes and endless piles of invoices. The through-line is an old-fashioned one: the most valuable AI is the AI that does the grunt work nobody wants to touch, and does it well enough that you don’t have to think about it.
The most visible part of this wave is the rise of personal AI assistants that live where people already communicate. Instinct, perhaps the hottest AI agent startup in Silicon Valley, reached a $10 billion valuation before even releasing its product to the public, and it is currently handling a reported $1 billion in annual transaction volume, mostly from travel-related requests. It lives inside iMessage and WhatsApp, not in a separate app, and it has its own email address, a private browser, and the ability to make phone calls, all of which allow it to book flights, track orders and run simple errands. Its CEO, Noah Shinn, is only 23 years old, and the company has raised $1 billion from Sequoia, Benchmark and Coatue. Meta has jumped into the game with Muse, a personal agent that launched in early September and crossed five million downloads in three weeks. Muse can book flights and send emails, but its breakout use case has been personal finance: early users used it to negotiate bills, chase refunds and cancel subscriptions, sometimes recovering thousands of dollars. Meta promoted the app aggressively across Facebook and Instagram, and it even introduced a Tamagotchi-style wearable called Muse Charm so people can interact with their agent away from their phones. OpenAI, not to be outdone, launched Dots, a set of “always on” agents that can juggle tasks in the background and connect with thousands of apps. There are also more careful takes on the personal assistant. Wajo’s assistant, Fo, is designed with a pragmatic twist: if the AI gets stuck on a phone call or a website verification, a trained human contractor quietly steps in to finish the job, so the user never sees a frozen app. And Known is using AI to do something even more human: dating. Known’s AI interviews users about their exes, preferences and ambitions, then makes one introduction at a time with a tailored explanation of why the match might work. When both people accept, Known acts as a first-date concierge, charging $15 a date and choosing the location. It only operates in the Bay Area and San Diego for now, but the company says 60% of its introductions are accepted and it has scheduled thousands of dates. These agents are not just chat interfaces; they are closer to personal assistants with a pulse, and they are beginning to feel like companions rather than tools.
Beyond personal assistants, a growing group of startups is selling AI agents as coworkers, employees and middle managers. Town, founded by Jean-Denis Greze, sells “townies” to small business owners, solo entrepreneurs and sales teams, focusing on work that lives in the inbox and dies without follow-up. For a restaurant owner, that might mean monitoring supplier price lists for increases; for a recruiter, it might mean following up with clients and compiling candidate reports. Unlike free consumer assistants, Town charges $50 to $200 per user per month, and it markets itself on what it does not keep: it doesn’t train on customer data and deletes most information after 15 days. Viktor is selling something even more ambitious. Its AI agents, which live in Slack and Microsoft Teams, are meant to be handed a business goal, not a list of prompts. You give Viktor a revenue target or a department to assist, and it wrangles the details, writes code, produces reports and schedules recurring tasks. The company claims it reached a $50 million annualized run rate in less than seven months. Andon Labs has pushed the idea further by trying to have AI agents run whole companies, including a store in San Francisco, a cafe in Sweden, and four radio stations. The store isn’t profitable, and the CEO admits that’s not the point; what matters is proving that agents can handle messy managerial work, from writing job postings to interviewing applicants. The company says it already has a waitlist of thousands of companies from more than forty countries, with proposals ranging from farms and hotels to a bank and even graveyard maintenance. Then there is TeamTangoAI, founded by a sixth-generation funeral director named Jimmy Lucas. He owns and operates 13 funeral homes in the Dallas-Fort Worth area, and he is using AI to answer overnight calls, texts and website chats that used to fall to exhausted staff or impersonal answering services. Each funeral home decides how much the AI can handle, but one rule is non-negotiable: if a family asks for a real person, the agent immediately complies. Lucas’ pitch is not about replacing funeral directors but giving them more time to care for families on their worst days. Element451 is bringing agents to higher education, where they act as AI admissions consultants, reading applications, nudging students to finish forms, answering questions about financial aid and even flagging fraud. It is unglamorous work, but it is exactly the kind of work that determines whether a student becomes a student or slips away.
Some of the most surprising uses of agentic AI are in places where software rarely goes. Atarraya, for example, is using AI agents to help workers run shrimp farms in shipping containers. The software maps shrimp biology, analyzes farm data and turns it into daily tasks, allowing the company to train former construction workers to become shrimp farmers in three weeks. The company is expanding in North America and the Middle East, and its success would be a powerful demonstration that AI can make messy biological operations manageable for people without specialized degrees. Salient is using agents to fight insurance companies on behalf of consumers, gathering claims files, researching vehicle values and negotiating with insurers after a car is totaled. It also handles credit-card disputes, debt collections and customer service, and it claims to be generating more than $40 million in annualized revenue. Stuut focuses on a problem every finance team knows: getting paid. Its agents chase invoices across email, text and phone calls, investigate billing disputes, determine whether deductions are justified and match payments to the right invoices. Stuut says it collected more than $21 million for one customer, and more than $3 billion has moved through its platform. Govra is working the other side of the counter, building AI agents that help government auditors. California’s tax department has already partnered with Govra on a pilot, and an official in Indiana sees the potential to conduct 20% to 30% more audits with its help. The CEO admits that more tax audits sounds scary, but his pitch is about fairness: if governments can quickly find tax cheats, the compliant majority wins. On the security front, XBOW is using AI to act as red teamers, probing customer applications, identifying vulnerabilities and building working exploits to show what an attacker could do. It serves more than 200 customers, including major banks and Fortune 500 companies, and it has found more than 14,000 exploits in real applications. And Nous Research, a Brooklyn open-source lab with a punk-rock aesthetic, is becoming a serious option for enterprises that want to run their own agents on their own servers, without handing their data to OpenAI or Anthropic. Its framework, Hermes, started as a tool for technical users who wanted to keep their data private, and now Fortune 500 companies are rolling it out to employees. It’s a reminder that in this market, trust and control are as important as raw capability.
Underneath all of these agents is a layer of infrastructure that is quietly making autonomous commerce possible. Browserbase isn’t trying to be another AI assistant; it wants to be the browser that agents use. As AI agents zip through the web, they need infrastructure to load sites, click buttons, enter information and complete transactions. Browserbase sells that layer to other agent companies, and it says its platform has already automated more than 2,000 years’ worth of browsing. Exa is building a search engine specifically for agents, so that before an agent can book a flight or draft a sales email, it can pull background information, trawl technical documentation and understand what it is looking at. More than 500,000 developers and 5,000 companies are building with Exa’s tools, and the company raised $250 million at a $2.2 billion valuation in May. The most delicate piece of infrastructure is payments. Letting an AI agent browse the web is easy; letting it spend your money is another matter. Stripe has become the default middleman for agentic commerce with its Link product. When a user approves a purchase, Link issues a virtual one-time-use credit card for the exact amount, so the agent completes the transaction without ever seeing the real credit card. Link also includes purchase protections for loss, theft and price drops. Meta’s Muse launched with Link, and buzzy agents like Instinct and Town use it too. Stripe said that agentic purchases using Link increased 38x in the past month. Even SpaceXAI’s Grok Bot, which started as an internal agent for SpaceX and now has a team of bots that collaborate on tasks and are managed by a “chief of staff” agent, relies on this kind of infrastructure. The point is not just that agents are getting smarter; it’s that the rails they run on are getting safer and more reliable, which is what finally lets people hand them real responsibility.
If there is a lesson in this list, it is that the future of AI agents is not one single breakthrough but a thousand small, disciplined acts of reliability. The agents that are getting traction are not always the ones with the most impressive demos; they are the ones that work well enough to be trusted with a phone call, a bank account, or a funeral home’s after-hours messages. And the best ones understand their limits. Wajo brings in human contractors when the model is stuck. TeamTangoAI insists on transferring to a real person when asked. Stripe uses one-time virtual cards precisely because it doesn’t fully trust agents with real financial details. Even the most aggressive players in this space are still experimenting, still running pilot programs, still waiting to see whether profitability follows trust. Andon Labs admits its AI-run store isn’t profitable, but it argues that the point is to prove that agents can handle real-world messiness. Govra is betting that as taxpayers use AI to navigate bureaucracy, governments will need agents of their own. And all of these companies are chasing a vast market: the businesses and individuals who don’t want a chatbot with a to-do list, but a colleague, an assistant, a fixer, a night-shift receptionist, a bill collector, a matchmaker, a shrimp-farming specialist. The promise of agentic AI, as this list makes clear, is not artificial general intelligence. It is something more immediate and more practical: the slow, unglamorous, but very real takeover of the work that, until now, only a human being could be trusted to do. And if that work is done well enough, the result is not an unemployed workforce but a world where people can spend more of their time on the tasks that matter, and let the agents handle the rest.
