What does the AI agent startup landscape look like today?

What actually counts as an AI agent startup today?

An AI agent startup today is a company whose main product can decide how to carry out multi-step work and then act through software, data or other tools with limited human supervision.

We need that cutoff because the word “agent” has spread across almost the entire AI industry. A chatbot that answers questions from company documents is usually still an assistant. An agent can take the next step itself: inspect an account, choose a tool, change a record, send something, run code, navigate a website or keep working until a task reaches a defined outcome.

That puts companies such as Sierra in customer service, Cognition in software engineering, Harvey in legal work, Hippocratic AI in healthcare and Arcade in secure agent execution inside the landscape we are studying. We would leave foundation-model companies such as OpenAI and Anthropic outside the startup map, even though they increasingly compete with these companies, because their businesses extend far beyond agents.

We would also leave out ordinary SaaS companies that have simply added an agent feature. The useful question is whether autonomous work sits at the center of what customers are buying.

Are companies actually letting AI agents work on their own yet?

Companies are using AI agents now, but most enterprises still keep them inside narrow workflows with clear limits on what they can do.

McKinsey’s latest research captures the gap well. Around 62% of companies are experimenting with agents, yet fewer than 10% have scaled agents inside any individual business function. ServiceNow’s 2026 Enterprise AI Maturity Index found something similar from a larger survey of 4,500 executives: 59% of organizations were already using agentic AI and another 30% were piloting it, while only 5% were redesigning work around the technology.

ServiceNow found zero surveyed companies with a genuinely cross-functional, self-improving agentic operating system. Most agents still handle pieces of work such as ticket triage, information gathering, coding tasks or customer requests rather than running an entire department.

The bottleneck is also shifting. McKinsey’s most recent work on agent adoption found that organizational change and siloed ways of working now rank ahead of technology infrastructure among the obstacles to scaling AI. Its researchers describe a rough 1:3:5 pattern among successful transformations: for every dollar spent on agent technology, companies may need three dollars of process redesign and five dollars of capability building and adoption.

Enterprise adoption is clearly real. Full autonomy is much further away than the number of “agent deployments” might suggest.

Which AI agent markets are actually working today?

The strongest AI agent markets today are customer service, software engineering, legal work, healthcare engagement and enterprise knowledge, where companies can give an agent a clear job and measure whether it completed that job properly.

That common structure is more important than the industry label. Customer-service teams can measure resolution rates and handling time. Engineering teams can inspect the code that an agent produces. Lawyers can review research, contracts and filings. Healthcare providers can measure whether patients answer calls, complete follow-ups or adhere to care plans.

General-purpose agents face a harder commercial problem because the product has to be reliable across many unrelated tasks. The more tightly a startup can define what success looks like, the easier it becomes to prove that the agent is worth paying for.

AI agent market	What the agent actually does	Commercial evidence today
Customer service	Resolves requests and changes customer records	Sierra, Decagon and Fin have reached large enterprise deployments
Software engineering	Writes, fixes, migrates and tests code	Cognition is already generating hundreds of millions in annualized revenue
Legal	Researches, drafts and executes legal workflows	Harvey is used across major law firms and corporate legal teams
Healthcare	Calls patients, follows up and coordinates care	Hippocratic AI has processed hundreds of millions of patient interactions
Enterprise knowledge	Finds company context and turns it into actions	Glean has grown past $300M in annualized recurring revenue

Why is customer service ahead of other AI agent markets?

Customer service is currently the clearest breakout AI agent market because companies already spend heavily on the work, requests arrive in huge volumes and outcomes can be measured immediately.

Sierra gives us unusually strong evidence. The company says its agents now serve more than 40% of the Fortune 50 and handle billions of customer interactions. At Singtel, Sierra reports resolution rates above 70% after a deployment that took less than ten weeks. Cigna went live in eight weeks and cut the time needed to authenticate a patient by 80%.

Decagon reached the same market from another direction. It added more than 100 global enterprise customers during its previous fiscal year, including Avis Budget Group, Block and Deutsche Telekom. Its pricing can be tied directly to conversations or successful resolutions, which pushes the product closer to paying for completed work than paying for another software seat.

The category is also expanding beyond one-off support conversations. Sierra recently launched Horizon for processes that can last days or weeks, such as collecting overdue payments, refinancing loans or settling insurance claims. Its new Plaid partnership allows agents to use bank information when those workflows require it.

Customer service is giving us an early picture of where agent software may go: longer-running tasks, more access to business systems and pricing linked more closely to outcomes.

Are coding agents becoming a real standalone software market?

Coding agents are already a large standalone software market, although independent startups face unusually aggressive competition from the model companies themselves.

Cognition is the strongest startup example we have today. The company said Devin and its wider product portfolio had reached about $492 million in annualized revenue when it raised more than $1 billion at a $26 billion post-money valuation. Enterprise usage of Devin had been growing roughly 50% month over month for six consecutive months.

The jobs companies give Devin are also becoming clearer. Cognition says customers such as Mercedes-Benz, NASA, Goldman Sachs and Santander use the product, while CEO Scott Wu has described migration work, software modernization and other long-tail engineering tasks as particularly useful workloads. These are expensive jobs that companies often postpone because human engineers would rather work on higher-value projects.

More recently, TechCrunch reported that Cognition was already discussing another financing at a valuation of at least $40 billion, with a possible $1 billion annualized revenue run rate forming part of the fundraising discussion. That number has not yet been announced as achieved revenue, so we should treat it differently from the confirmed $492 million figure. Still, the fact that the conversation has moved that far in only a few months shows how quickly this category is growing.

The pressure comes from Claude Code, OpenAI Codex and Google’s coding products. Coding may become one of the biggest agent markets while also becoming one of the toughest places for an independent startup to hold its ground.

Are vertical AI agents beating general-purpose agents?

Vertical AI agents currently have the stronger business case because specializing in one kind of work gives startups more ways to differentiate beyond whichever foundation model they use.

Harvey is the clearest example. The legal AI company now generates roughly $350 million in annual recurring revenue according to recent reporting from The Wall Street Journal. Even more interesting is usage: monthly token consumption increased from around 1 trillion to 14.5 trillion in six months. That is a fourteenfold jump in actual product consumption inside a customer base dominated by lawyers and legal teams.

Harvey has lately pushed further into the legal stack. Harvey II adds persistent memory around how individual lawyers work, while Harvey Tenet gives the company a legal model built from an open-weight foundation model and specialized for legal tasks. That reduces dependence on paying frontier-model providers for every piece of intelligence while giving Harvey more control over cost and performance.

Healthcare creates a similar specialization advantage. Hippocratic AI’s current Polaris platform says it has been built on more than 200 million patient interactions. The company develops separate healthcare agents and orchestrators for providers, insurers and life-sciences companies rather than asking one universal agent to handle every medical workflow.

EliseAI gives us a third example outside the usual legal-and-healthcare pair. The company passed $100 million in ARR while automating workflows for housing operators and has lately been discussing a financing that could value it around $3.7 billion. Its agents handle apartment inquiries, scheduling, maintenance communication and related administrative work, with healthcare becoming an additional market.

We would put more weight on startups that understand a difficult profession than on startups whose main pitch is that their agent can do almost anything. The specialized product simply has more places to build an edge.

If you want more recent data on this point, please see our latest agentic AI market report.

Is company data becoming a bigger AI agent moat than the model itself?

Company data and context are becoming some of the strongest moats in enterprise AI agents because even a brilliant model is much less useful when it cannot understand how a particular company actually works.

Glean is the best current example. The company reached $300 million in annualized recurring revenue only 15 months after crossing $100 million, while its Fortune 500 customer count nearly doubled year over year. More than 85% of customers reportedly use Glean across at least five departments.

Glean had spent years indexing enterprise information before the agent boom arrived. Its software understands permissions, relationships between employees and documents, internal applications and company-specific knowledge. Glean Agents can now use that existing context instead of reconstructing it for every task.

That history changes the competitive equation. A customer can switch from one frontier model to another relatively quickly. Rebuilding a trusted map of who is allowed to see what, where the company’s information sits and how different systems connect is much harder.

These days, that context layer may be worth more than having slightly better raw reasoning.

Can general-purpose AI agent startups survive OpenAI, Anthropic and Google?

General-purpose AI agent startups can still build very large businesses, but they sit in the most exposed part of the startup landscape because their ambition overlaps directly with OpenAI, Anthropic, Google, Meta and Amazon.

Adept showed the danger early. The startup raised hundreds of millions of dollars around the idea of a general agent that could operate software. Building both its own large models and the agent product proved extremely capital-intensive, and Amazon eventually hired several founders and team members while licensing Adept’s technology. Adept itself shifted toward a narrower agent-infrastructure strategy.

Manus shows the upside as well as the strategic instability. The general-purpose agent went from launch to more than $100 million in ARR remarkably quickly. Sacra now estimates annualized revenue around $450 million. Meta agreed to buy the company for roughly $2 billion, but Chinese authorities later forced the transaction to unwind, leaving Manus independent again and exploring new financing options.

The commercial demand is clearly there. The strategic position remains uncomfortable. A general agent wants to become the place where users delegate almost any digital task, which puts the startup directly in the path of companies that already control the model, operating system, browser, cloud or consumer distribution.

A founder can still win here. The bar is simply much higher than building a better autonomous demo.

Are OpenAI and Anthropic making the basic AI agent stack a commodity?

OpenAI, Anthropic and other model providers are steadily commoditizing the basic machinery needed to build AI agents, which makes generic orchestration a much weaker startup moat today.

OpenAI’s Agents SDK now includes handoffs between agents, guardrails, tracing, approvals, long-running state and controlled sandbox environments where agents can inspect files, execute commands and edit code. A startup previously had to build much of that infrastructure itself.

The direction has become even clearer lately. OpenAI is winding down its hosted Agent Builder and Evals products and directing developers toward the code-first Agents SDK. That concentrates more of the durable agent plumbing around a foundation-model ecosystem.

Anthropic has put similar pressure on integration startups through MCP. Developers can increasingly connect Claude and other compatible models to tools through a standard protocol rather than writing proprietary integrations from scratch.

That pressure is unlikely to reverse. Model companies want developers spending less time constructing an agent loop and more time consuming their models. Every technical obstacle they remove forces startups that were charging mainly for that obstacle to move somewhere harder.

If you want more recent data on this point, please see our latest agentic AI market report.

Are MCP and A2A killing the old AI agent integration moat?

MCP and A2A are quickly weakening the idea that simply connecting an AI agent to lots of applications can support a durable startup.

MCP has become far larger than an Anthropic-specific protocol. When the Model Context Protocol moved into the Linux Foundation’s Agentic AI Foundation, more than 10,000 published MCP servers already existed across developer tools, business applications and enterprise deployments.

Google’s Agent2Agent protocol is following a similar path for communication between agents. A2A launched with support from more than 50 organizations; the Linux Foundation now says more than 150 organizations support the standard, including AWS, Microsoft, Salesforce, SAP, ServiceNow and Cisco. Production-ready SDKs cover five major programming languages.

The institutional adoption has accelerated further. The Agentic AI Foundation recently reached 247 member organizations after adding another 57 members, including Visa, Wells Fargo and Alibaba.

MCP handles how agents reach tools and data. A2A handles how agents find and communicate with other agents. Once those connections become standard, maintaining a large proprietary catalogue of integrations becomes much less impressive.

There will still be money in integrations, especially where authentication and reliability are difficult. The value is moving toward making those connections safe and dependable rather than merely making them exist.

Where are the best AI agent infrastructure startups forming now?

The most interesting AI agent infrastructure startups today are moving into execution, browser access, identity and security because production agents still struggle with those problems even when the model can reason correctly.

Arcade.dev is a good example of the shift. The company raised $60 million for its secure action layer after tool-call volume increased 25-fold in six months. Arcade sits between an agent and the systems it wants to modify, handling delegated authorization, permissions, auditability and governance. Its founders also helped write the MCP authorization specification.

NewCore has emerged around an even more basic problem: identity. The company raised $66 million to build an identity platform that treats AI agents as separate identities with their own permissions, lifecycle and revocation controls. NewCore currently has fewer than ten customers and more than ten design partners, so this is still an early market rather than proven infrastructure at scale.

Browserbase is much further along on usage. Its platform now handles more than 35 million browser sessions per month for customers including Ramp, Shopify and Lovable. Airtable alone has run more than 300,000 Hyperagent sessions on Browserbase since launching that product. Browserbase recently turned the underlying infrastructure into a managed browser-agent product that can operate websites without companies maintaining a separate automation script for every site.

These companies sit below the glamorous part of the agent. They solve what happens after the model decides what it wants to do: log in, obtain permission, operate a messy website, preserve state, execute safely and leave a record of what happened.

That layer looks increasingly valuable.

Are multi-agent systems actually happening yet?

Multi-agent systems are starting to appear in real products, but most companies are still far from running large autonomous teams of AI agents.

The technical direction is easy to see. OpenAI’s current agent tooling supports handoffs and agent-as-tool delegation. A2A has reached production use specifically because companies want agents built on different systems to coordinate with each other.

Hippocratic AI is already applying that design to healthcare through what it calls Agentic Orchestrators. Instead of asking one healthcare agent to do everything, an orchestrator can coordinate specialized agents around a broader objective such as reducing readmissions, closing gaps in patient follow-up or keeping a clinical trial on track.

Enterprise surveys remain much more conservative. ServiceNow found that companies mostly use disconnected agents or assistants, while genuinely cross-functional agent systems were effectively absent from its sample.

For now, multi-agent systems are more advanced as an architecture than as a normal enterprise operating model. That gap should close gradually as companies become comfortable letting individual agents handle larger pieces of work.

Will big software companies crush AI agent startups or buy them?

Big software companies are increasingly doing both: building their own agents while buying startups that have already solved a valuable part of the problem.

Salesforce makes the pattern particularly visible. Agentforce has now reached $1.2 billion in ARR, up 205% year over year, according to Salesforce’s latest reported quarter. The company also says Agentforce and Slack have delivered 3.8 billion “agentic work units,” its internal measure of work performed by agents.

Yet Salesforce has kept buying outside technology. It acquired Convergence.ai for agents that can navigate software interfaces and Spindle AI for agentic analytics. Most recently, Salesforce agreed to pay about $3.6 billion for Fin, formerly Intercom, whose customer-service agent already has an established customer base and its own support-focused AI model.

The Fin deal is especially revealing because Salesforce already had a large customer-service agent product. Paying billions for another one suggests that distribution alone does not instantly reproduce several years of product learning, customer deployment and domain-specific technology.

For startups, incumbents create two risks at once. They can bundle competing agents into software that customers already own, and they can buy the strongest independent companies before those businesses become larger threats. That makes acquisition part of the competitive structure of this market, not just an occasional exit.

If you want more recent data on this point, please see our latest agentic AI market report.

Is AI agent startup funding still accelerating?

Investors are putting much more money into AI agents, but the current funding market is splitting between a huge number of young companies and a small group of extremely expensive leaders.

Crunchbase’s current agentic-AI database lists 188 companies with early-stage venture funding and 540 funding rounds worth $8.6 billion cumulatively. Its broader dataset contains hundreds of agentic-AI companies founded within the past year alone. The exact totals depend heavily on how broadly we define “agentic AI,” so we would avoid pretending that one database gives us a precise market size.

The large rounds are easier to verify. Cognition raised more than $1 billion. Sierra raised $950 million. Decagon raised $250 million. Harvey raised $200 million. NewCore emerged with $66 million, while Arcade raised $60 million.

Those six transactions alone add up to more than $2.5 billion. Cognition and Sierra account for more than $1.95 billion of that amount.

Startup formation is spreading everywhere while serious capital is concentrating very quickly around companies that already have customers, usage or a plausible infrastructure choke point.

AI agent company	Recent disclosed financing	What investors are backing
Cognition	>$1B	Coding agents at large commercial scale
Sierra	$950M	Customer-service agents
Decagon	$250M	AI customer concierge
Harvey	$200M	Legal agents
NewCore	$66M	Agent identity and security
Arcade.dev	$60M	Secure agent actions and authorization

Are the biggest AI agent startups already real businesses?

Several leading AI agent startups are already producing hundreds of millions of dollars in annualized revenue, which puts the top of this market well beyond the usual pre-revenue AI hype cycle.

Glean has officially reported more than $300 million in ARR. Recent reporting puts Harvey around $350 million. Cognition reported a $492 million annualized revenue run rate at its last financing. Third-party private-market estimates put Sierra around the $200 million range, although Sierra itself has not publicly confirmed that figure.

These numbers require some care. Consumption-based AI companies often describe annualized usage as ARR even though usage can move up or down more quickly than a traditional SaaS subscription. Glean, for example, offers consumption-based and hybrid pricing, so even its $300 million headline does not behave exactly like old-fashioned recurring software revenue.

Still, the scale is hard to dismiss. A handful of agent companies have gone from almost no revenue to several hundred million dollars within two or three years, and some are doing it while selling to large regulated enterprises.

Company	Main market	Latest public or reported revenue level	Latest closed valuation
Cognition	Coding	~$492M annualized run rate	$26B
Harvey	Legal	~$350M ARR	$11B
Glean	Enterprise knowledge and agents	>$300M annualized recurring revenue	$7.2B
Sierra	Customer service	~$200M estimated ARR	>$15B

Are AI agent startup valuations getting out of control?

Some AI agent valuations are clearly pricing years of exceptional growth in advance, and the biggest risk today is that revenue growth slows before those expectations come down.

Sierra’s last round valued the company above $15 billion. Using the roughly $200 million third-party ARR estimate gives a multiple above 75 times annual revenue. Cognition’s last closed $26 billion valuation against its disclosed $492 million annualized revenue was above 50 times.

Cognition now shows how quickly the denominator can move. As seen above, the company disclosed $492 million in annualized revenue only recently, yet another fundraising discussion is reportedly being framed around reaching roughly $1 billion. If that happens quickly, an apparently extreme valuation multiple compresses just as quickly.

Harvey offers a somewhat less aggressive comparison. Roughly $350 million of ARR against its last $11 billion valuation works out to around 31 times revenue. Glean’s $7.2 billion valuation against more than $300 million is below 24 times, although its mixed consumption model makes a simple SaaS comparison imperfect.

These are still very expensive companies. Investors are betting that agent startups can capture spending that previously went to human labor, outsourcing and services in addition to normal software budgets.

That could create businesses much larger than classic SaaS categories. It also gives investors plenty of room to be spectacularly wrong.

If you want more recent data on this point, please see our latest agentic AI market report.

What separates a durable AI agent startup from a wrapper today?

A durable AI agent startup today needs to own a part of the workflow that stays valuable even when the underlying models become cheaper and much better.

We can already see several versions of that advantage. Glean owns enterprise context and permissions. Harvey is accumulating legal workflows, memory and specialized models. Arcade controls how agents receive permission to act. Browserbase runs the browser infrastructure agents need to reach software that has no clean API. Sierra is pushing deeper into long-running customer workflows that connect conversations with payments, claims and financial data.

The weakest position is much easier to identify now than it was two years ago. Generic prompting, basic orchestration and large lists of simple connectors are all becoming cheaper to reproduce. MCP, A2A and foundation-model SDKs accelerate that process.

Distribution also matters more than many technical agent maps suggest. An agent that sits directly inside a customer-service operation, law firm, hospital or engineering team learns how that organization works and becomes harder to remove. An impressive horizontal demo has to earn that position from scratch.

Over time, the strongest companies will probably stop being described primarily as “AI agent startups.” Harvey will look increasingly like legal software. Sierra and Decagon will compete for customer operations. Hippocratic AI will sit inside healthcare delivery. Arcade and NewCore will resemble security infrastructure.

“Agent” describes how the software works. The durable business comes from the job the software owns.

So what does the AI agent startup landscape look like today?

The AI agent startup landscape today is growing extremely fast, but the market is already moving away from generic agents and toward companies that own expensive, specific pieces of real work.

Enterprise demand has moved beyond experimentation, although genuinely autonomous workflows remain rare. Customer service is the most mature application market so far. Coding has produced the fastest-growing independent agent company we can identify. Legal and healthcare show why domain expertise can become a serious moat. Glean shows how valuable enterprise context can be. Arcade, NewCore and Browserbase show a second large opportunity forming underneath applications around execution, identity, security and access.

Meanwhile, the generic pieces of the technology stack are getting cheaper. OpenAI is packaging more agent infrastructure into its SDK. MCP is standardizing connections to tools. A2A is standardizing communication between agents. The Agentic AI Foundation now has hundreds of member organizations pushing those standards forward.

That changes what we should look for in startups. Being able to build an agent is becoming less impressive every quarter. Getting an agent trusted inside a company, giving it the context and permissions it needs, making it reliable enough to complete valuable work and capturing part of the economic value of that work are becoming much more important.

Our conclusion is fairly sharp. AI agents are becoming a major software market, but “AI agent startup” will probably prove to be a temporary category. The strongest companies are already turning into customer-service platforms, coding companies, legal systems, healthcare infrastructure and security businesses whose products happen to be powered by agents.

That is where we would expect most of the durable value to end up.

If you want more recent data on this point, please see our latest agentic AI market report.

Who is the author of this content?

NEW MARKET PITCH TEAM

We track new markets so founders and investors can move faster

We build living "market pitch" documents for emerging markets: AI, synthetic biology, new proteins, and more. Instead of outdated PDFs or hallucinated LLM answers, our clients get a clean, visual, always-updated view of what's really happening: key players, deals, regulations, and signals that matter. Learn more about us.