AI Agents: what’s changing now?
In our agentic AI market deck, you will find everything you need to understand the market
SUMMARY
AI Agents: what’s changing now? AI agents are becoming a real execution layer for digital work: they can increasingly pursue multi-step tasks, use tools, change external systems and hand back completed work rather than just answers.
Coding is the first agent market that has clearly reached mass commercial scale. Claude Code, Codex and Devin combine strong revenue or usage with a work environment where outputs can be tested, failures can be reversed and valuable tasks can be delegated safely.
The more interesting behavioral shift is not people walking away from AI for an entire day. It is people assigning work that would take humans hours, running several agent jobs in parallel, and moving into a supervisor role where they define outcomes and review results.
Customer service now looks like the strongest large market after coding. Adoption is rising quickly, resolution rates are high enough to matter economically, and vendors are beginning to charge for completed outcomes instead of simply selling another software seat.
Enterprise adoption is real but still shallow. Many organizations are experimenting, a meaningful minority are scaling agents somewhere, and companies that get useful deployments working tend to create more agents — yet broad rollout across many functions remains uncommon.
The best agent workloads share a practical trait: success is easy to check and mistakes can be contained. Computer use expands the addressable market by letting agents operate old software without APIs, but structured integrations remain cleaner whenever they exist.
A technical stack is forming around the market. MCP is becoming the common layer for connecting agents to tools and data, while A2A is emerging for agent-to-agent communication across systems and organizational boundaries.
Independent agent startups can still become very large businesses, but they need more than access to a frontier model. Durable value is moving into integrations, workflow logic, permissions, testing, recovery systems, context and the accumulated operating data around the job.
Reliability and security are now tightly linked. As agents gain permission to act, a small reasoning error or malicious instruction can propagate into real systems, which is why observability, least-privilege access, approvals, sandboxes, identities and audit trails are becoming part of the product rather than optional enterprise extras.
The labor shift is happening at the task level faster than at the job level. At the same time, software economics are starting to move from seats and prompts toward resolved cases, completed jobs and measurable work — a harder standard, but also a much more meaningful one.
AI Agents: what’s changing now?
What does “AI agent” actually mean now?
AI agents now increasingly mean software that can pursue a task, use tools and take actions across several steps with limited supervision.
That is narrower than the way “agent” is often used in marketing. A chatbot that answers a question from company documents may use an advanced model, but there is little agentic behavior if the system never decides what to do next or changes anything outside the conversation.
The clearer examples are products such as Claude Code, Codex and Devin. They can inspect repositories, modify files, run commands, test their work and continue after failures. Enterprise products are moving in the same direction. Microsoft Copilot Studio can trigger agents from business events and let them work across applications, while customer-service agents increasingly retrieve account data, apply company policies and execute actions for customers.
Full autonomy is still rare. Most useful AI agents currently work inside boundaries set by humans: approved tools, specific permissions, spending limits, escalation rules or predefined environments. An agent allowed to issue a refund below a certain amount can still be useful even though it cannot rewrite the refund policy.
The real break from the chatbot era is simple: AI can increasingly touch the systems where work happens.
Why are AI agents suddenly getting so much better?
AI agents are getting much better because models can now stay useful across longer chains of work, while the software around them has become far better at keeping those chains under control.
Model intelligence obviously helps, but reliability across many steps is the bigger issue. Suppose an agent makes the correct decision 90% of the time at each stage. Across ten independent decisions, its chance of getting the entire sequence right falls to roughly 35%. Small improvements in individual steps can therefore make a surprisingly large difference to whether a long workflow works at all.
METR has been measuring this through the length of tasks frontier agents can complete with 50% reliability, using the time a skilled human would need for the same task. Its research finds that this “task horizon” has historically doubled about every seven months. The researchers also found that agents are much more reliable on short tasks than long ones, which fits what people experience in practice: knowing how to perform every individual step does not guarantee that an AI can string all of them together correctly.
The systems surrounding the models have improved at the same time. Agents increasingly work with explicit tools, isolated environments, stored context, retries, tests and approval gates. Microsoft recently said an upgraded Copilot Studio orchestration layer improved its evaluation performance by around 20% while cutting net token use by 50%.
This market map, featured in our agentic AI market deck, highlights top companies and startups in the agentic AI market
Are coding agents the first real mass-market AI agents?
Yes. Coding agents currently provide the strongest evidence that people will use and pay heavily for AI that actually performs work.
Claude Code has already passed $2.5 billion in run-rate revenue, according to Anthropic, after more than doubling since the start of the year. Weekly active users doubled over the same period, business subscriptions quadrupled, and enterprise customers now account for more than half of Claude Code revenue. Anthropic also cited an external estimate suggesting Claude Code was responsible for roughly 4% of public GitHub commits worldwide, twice the share measured one month earlier.
OpenAI says Codex is used by more than four million people each week. The company has also seen developers move from autocomplete toward delegating entire jobs such as debugging, refactoring, writing tests and implementing features.
The independent market is growing alongside the model companies. Cognition disclosed this week that Devin's annualized run-rate revenue had risen from $492 million to almost $900 million since its previous financing round. That is roughly an 83% increase from an already large base in only a few months.
Software engineering gives agents almost ideal conditions. The work is digital, outputs can be tested, failures can often be reversed, repositories contain useful context and a completed task can be worth far more than the computing cost. An agent can also work inside a sandbox instead of receiving unrestricted access to the real world.
Coding therefore looks much less like a niche early-adopter market these days. It is the first place where the agent model has genuinely broken through.
| Coding-agent evidence | Current scale |
|---|---|
| Claude Code | More than $2.5B run-rate revenue |
| Codex | More than 4M weekly users |
| Devin | Nearly $900M annualized run-rate revenue |
| Claude Code business subscriptions | Up 4x since the start of the year |
If you want more recent data on this point, please see our latest agentic AI market report.
Are people really leaving AI agents alone for hours?
People are currently delegating much longer jobs to AI agents, although the more interesting behavior is supervising several agents at once rather than disappearing for an entire day.
OpenAI studied how Codex was being used and found that 80.6% of sampled individual users had assigned at least one task estimated to represent more than 30 minutes of human work. Some 70.2% had delegated a task exceeding one hour, while 25.6% had tried one estimated above eight hours.
Those figures refer to estimated human-equivalent work, so they should not be read as proof that an agent ran independently for eight straight hours. That’s a big caveat. They still show a behavioral change: people are becoming comfortable describing a substantial outcome and returning later to inspect what happened.
Parallelism pushes this further. OpenAI found that some heavy internal Codex users were generating more than 60 hours of agent turns per day because several tasks could run simultaneously. A human obviously cannot personally work for 60 hours in one day, but one person can launch several jobs, review the results and decide what happens next.
As this chart shows, and as featured in our agentic AI market deck, search interest in AI agents has been rising rapidly
Are enterprise AI agents actually moving beyond pilots?
Yes, although company-wide AI-agent deployment is still much less common than the headlines suggest.
McKinsey's broad survey found that 62% of respondents said their organizations were at least experimenting with AI agents. Only 23% said an agentic system was being scaled somewhere in the business, and fewer than 10% reported scaled agent use in any individual function.
A survey from LangChain reaches a more technical audience and therefore shows a market further along. Among more than 1,300 respondents, 57.3% said their organizations already had agents in production, compared with 51% in the previous survey. Another 30.4% were actively developing agents with plans for production. Among companies with more than 10,000 employees, 67% reported agents already running in production.
Salesforce's latest usage data adds something the surveys cannot: behavior after deployment. Among businesses that consistently used Agentforce, the average number of activated agents nearly tripled over roughly a year, while the average time required to create one fell 53%. Salesforce also developed an “Agentic Work Unit” to count tasks actually completed by agents and found output growing at a 15% compound monthly rate by the end of its measurement period.
Deep deployment across many functions remains unusual, but companies that have agents working successfully tend to create more of them and give them more work.
Is customer service becoming the second big AI-agent market?
Yes. Customer service currently looks like the clearest large-scale agent market after software engineering.
A Salesforce survey of 3,075 service professionals found that AI-agent adoption in customer-service organizations rose from 39% to 66% in a year. That is a 27-point increase, or roughly 1.7 times the previous adoption rate. Among organizations already using service agents, 70% said they saw measurable value within 60 days.
The most interesting result was where that value appeared. Respondents ranked customer satisfaction as the KPI that improved most after introducing agents, ahead of representative productivity, average handling time, customer retention and first-response time. That weakens the idea that service agents are useful only because they make support cheaper.
Intercom offers another useful clue in the way it measures Fin. Its main automation rate counts the share of all customer conversations that Fin resolves without a human teammate having to step in. Intercom says Fin currently resolves an average of 76% of the customer queries it handles across its customer base, although that company-reported figure naturally reflects its own installed base.
Intercom has also moved toward outcome pricing. A normal Fin resolution costs $0.99, while sales qualification can be charged as a completed outcome rather than as a software seat. It sounds like a small pricing detail, but the implication is bigger: agent software can increasingly be sold against work performed rather than access to the software.
| Customer-service evidence | What we see now |
|---|---|
| Organizations reporting AI-agent adoption | 39% → 66% in Salesforce survey |
| Users seeing measurable value within 60 days | 70% |
| Intercom's reported average Fin resolution rate | 76% |
| Emerging pricing model | Payment for completed outcomes |
If you want more recent data on this point, please see our latest agentic AI market report.
This chart, included in our agentic AI market deck, illustrates yearly VC funding for agentic AI startups
What kind of work can AI agents actually operate on their own today?
AI agents currently work best on digital tasks that are structured enough to judge, while computer use is widening the number of systems they can operate without a custom integration.
Microsoft's computer-using agents in Copilot Studio are now generally available. They can navigate websites and desktop applications through a graphical interface, click buttons, fill fields and work through systems where no usable API exists. Microsoft specifically points to jobs such as invoice processing, data entry and extracting information from legacy applications.
APIs are still preferable whenever they are available. A structured connection is usually faster, easier to permission and less fragile than asking a model to inspect pixels and navigate a user interface. Computer use matters for the large amount of business software that was never designed to be operated programmatically.
The strongest agent tasks tend to share a few characteristics. Coding works well because agents can run tests and inspect whether their changes work. Customer service works because companies already have policies, historical conversations and escalation paths. IT operations, document processing and internal administrative work can also offer clear ways to verify whether the result is correct.
The harder cases involve hidden context, irreversible consequences or success criteria that are difficult to verify. An agent can prepare an expense report more safely than decide a company's acquisition strategy. It can extract information from invoices more easily than decide which employee should be fired.
Microsoft's production design reflects those limits. Companies can restrict which websites or applications an agent accesses, store credentials through managed systems, add human approval checkpoints and inspect what the agent saw and clicked.
Computer use is expanding what agents can touch, but the best workloads today are still the ones where success is easy to check and failures can be contained.
Is MCP really becoming the standard way AI agents connect to tools?
Yes. MCP currently has the strongest claim to becoming the common connection layer between AI agents and outside tools or data.
Anthropic released the Model Context Protocol as an open standard, and adoption quickly spread beyond Claude. OpenAI, Google, Microsoft and other major AI platforms now support MCP in parts of their ecosystems, alongside a fast-growing collection of developer tools and business applications.
The protocol has also moved away from single-company governance. Anthropic contributed MCP to the Agentic AI Foundation under the Linux Foundation, alongside founding contributions and support from several companies that compete directly in AI.
An agent becomes much more useful once it can search a database, retrieve documents, create a ticket, run a development tool or interact with a business application. Building every one of those connections separately for every AI product creates huge duplication.
MCP gives tool providers a common interface to build against. It will not eliminate proprietary connectors, and it does not solve permissions or security by itself, but it cuts one of the biggest pieces of integration friction around agents.
This chart, included in our agentic AI market deck, shows how Cognition is positioned in agentic AI
Are multi-agent systems becoming useful or are “agent swarms” mostly hype?
Multi-agent systems are becoming useful in specific situations, while the grander idea of autonomous agent swarms is still running ahead of the evidence.
The clearest progress is interoperability. Google's Agent2Agent protocol, now hosted by the Linux Foundation, passed support from 150 organizations within its first year. Google, Microsoft and AWS have integrated the protocol, and the foundation reports production use across areas including financial services, insurance, supply chains and IT operations.
A2A solves a different problem from MCP. MCP generally helps an agent access tools and information. A2A allows separately built agents to discover one another, communicate and coordinate tasks.
That becomes genuinely useful when organizational boundaries are involved. A company's purchasing agent might need to interact with a supplier's agent without either company handing over its internal systems. An IT agent could similarly ask a security agent to investigate something while each remains under separate policies and ownership.
Within one application, though, adding more agents can easily create unnecessary complexity. Each extra agent adds communication, context, permissions and another place for errors to appear. Many workflows work better with one capable agent and several well-designed tools.
If you want more recent data on this point, please see our latest agentic AI market report.
Are AI-agent startups becoming real businesses or just getting crazy valuations?
Some AI-agent startups are already growing into very large businesses, and investors are pricing them as if that growth can continue at an extraordinary pace.
Cognition gives us the freshest and clearest example. The company has just raised more than $2 billion at a $48 billion valuation, up from a $26 billion valuation at its previous financing. Over roughly the same period, Cognition says annualized run-rate revenue increased from $492 million to almost $900 million.
That means the revenue measure increased by around 83% while the valuation increased by about 85%. The close match is interesting: investors did not simply double the multiple while revenue stood still. They effectively repriced Cognition upward at roughly the same speed as its reported business run rate expanded.
The revenue number still needs careful interpretation. Cognition calls it annualized run-rate revenue rather than twelve-month revenue, and the company has not publicly explained all the mechanics behind the calculation. For a fast-growing usage-based AI product, annualizing a recent month can make the business look much larger than historical revenue would.
Costs matter too. The Information reported that Cognition's cash burn could approach $800 million this year because developing and operating coding agents requires expensive computing infrastructure. A company can therefore have spectacular top-line growth without resembling the high-margin SaaS businesses that traditional revenue multiples were built around.
Still, nearly $900 million of annualized revenue for a company founded in 2024 is hard to dismiss as speculative enthusiasm.
This chart, included in our agentic AI market deck, illustrates yearly funding for agentic AI startups
Are model companies going to crush independent AI-agent startups?
No clear winner has emerged. Independent AI-agent companies can still build valuable businesses, but surviving increasingly depends on owning much more than a clever prompt on top of somebody else's model.
Anthropic now competes directly in coding through Claude Code. OpenAI does the same with Codex. Microsoft, Google and Salesforce are building their own agent products and orchestration layers. A startup operating directly above frontier models therefore has to assume that its supplier may eventually enter the same product category.
At the same time, companies building agents increasingly use several model families. LangChain's latest survey found multi-model development had become normal among respondents, with OpenAI, Anthropic, Google and open-source systems all seeing substantial use. Fine-tuning was much less widespread.
That makes the surrounding system more important. Good agent products accumulate integrations, workflow logic, context, testing, permissions, recovery mechanisms and data from previous jobs. They can also route different work to different models as capabilities and prices change.
We can see this idea in how Cognition describes itself today: Devin can trigger work from systems such as Slack, GitHub and Linear, while Cognition positions its platform as independent enough to select models according to the task. The customer is buying the software-engineering workflow around the models as much as access to any specific model.
Foundation-model companies have enormous advantages in capability, compute and distribution. Independent agent startups still have room when they understand a particular job much better and can turn model intelligence into reliable execution.
If you want more recent data on this point, please see our latest agentic AI market report.
Has reliability stopped being the biggest AI-agent problem?
No. Reliability is still the biggest practical problem for AI agents in production, even as raw capabilities improve quickly.
LangChain's latest survey puts the issue plainly. Thirty-two percent of respondents named quality as a top barrier to deploying agents, making it the leading concern. Cost had become less prominent than in the previous survey.
The way companies are responding tells us even more. Nearly 89% of respondents had implemented observability for their agents, while only 52% had implemented evaluations. Organizations are racing to see what their agents actually did because unexpected behavior remains common enough to demand close inspection.
Long tasks make the problem harder. One small mistake can poison everything that follows. The model might call the wrong tool, interpret stale information as current, fail to notice that an application changed state or confidently continue after an earlier assumption became false.
Production systems therefore surround agents with checks that often receive less attention than the model itself. Tests verify output. Retries handle recoverable failures. Tool permissions restrict actions. Sandboxes contain mistakes. Human approvals appear before sensitive decisions.
This chart, included in our agentic AI market deck, compares the main business model options for autonomous AI agent platforms
Are AI-agent security problems getting worse as agents become more capable?
Yes. AI-agent security is becoming harder because agents can increasingly turn a bad instruction into an actual action inside company systems.
Prompt injection is the obvious example. A chatbot that reads malicious text may produce a bad response. An agent that reads malicious text while connected to email, cloud storage, internal databases and business applications has a much larger set of things it could potentially do wrong.
NIST has been focusing specifically on this problem. Its work on software-agent identity and authorization asks how companies should identify agents, authenticate them, enforce least-privilege access, record their actions and prove which human gave them authority. It also addresses direct and indirect prompt injection.
OWASP has pushed agent security further into its own category. Its recent work includes an Agent Control Standard alongside security guidance for agentic applications, covering problems such as goal manipulation, tool misuse, excessive privileges and unsafe execution.
The identity question will become especially important. A purchasing agent should not need every permission possessed by the employee who asked it to order office equipment. Companies need a way to say which agent acted, on whose behalf, with which permissions and within what limits.
Microsoft already exposes parts of this model in Copilot Studio through managed credentials, application restrictions, approval checkpoints and run histories.
Are AI agents actually replacing workers now?
AI agents are currently taking over meaningful pieces of jobs, while evidence of broad replacement of entire occupations remains much thinner.
Software engineering again gives us the clearest view. Anthropic studied interactions with Claude Code and classified 79% as automation, compared with 49% for ordinary Claude.ai usage. Coding-agent users were much more likely to hand work over to the AI instead of merely asking it for advice.
Cisco provides a company-level example. According to data published by OpenAI, Codex produced more than 95% of new code for one AI-feature deployment, increased defect-resolution throughput by roughly 10 to 15 times and saved more than 1,500 engineering hours per month. Engineers still owned the systems, reviewed work and handled harder decisions, but a meaningful amount of execution had moved to the agent.
Agent usage is also spreading beyond engineers. OpenAI's enterprise data shows that since February, weekly active Codex users increased 108 times in legal, 41 times in sales, 41 times in recruiting and 26 times in marketing, compared with five times in engineering. Those huge multiples partly come from tiny starting bases, so they should not be mistaken for evidence that lawyers or recruiters suddenly use Codex more than developers. They do show that agent-style work is escaping its original technical niche.
What we can say with confidence today is narrower: tasks are shifting faster than whole jobs. Employees increasingly define the job, review the output, deal with exceptions and keep responsibility for the final decision while agents do more of the execution in between.
This chart, featured in our agentic AI market deck, shows the share of revenue generated by each customer segment in the agentic AI market
What are companies measuring now that AI agents actually do work?
Companies are starting to measure completed work rather than simply counting AI usage, one of the clearest signs that the agent market is maturing.
Customer service has moved furthest. Intercom's automation rate measures what percentage of all conversations Fin fully resolves without a human stepping in. Salesforce's Agentic Work Unit similarly counts a task completed by an agent. It’s a slightly awkward metric, but a useful one: both try to answer a better question than “How often did somebody use AI?”
The same logic is visible in coding. Counting prompts says very little about productivity when one Codex task might represent a small bug fix while another represents hours of human-equivalent work. OpenAI has therefore started examining estimated task duration, parallel agent work and the proportion of enterprise output produced through agentic workflows.
Pricing is following measurement. Intercom charges for Fin outcomes. AI coding products increasingly mix subscriptions with usage or compute limits. If this continues, agent software could pull parts of the software market away from the classic “pay per employee per month” model.
That creates a tougher standard for vendors. A company can give every employee an AI assistant and report impressive adoption numbers. Once buyers pay for resolved cases, completed jobs or successful workflows, the vendor has to prove that useful work actually happened.
What is actually changing with AI agents now?
AI agents are currently becoming a real execution layer for digital work, with coding already at commercial scale and customer service moving quickly in the same direction.
Several independent pieces of evidence now line up. Claude Code has passed $2.5 billion in run-rate revenue. More than four million people use Codex weekly. Salesforce found service-agent adoption rising from 39% to 66%. LangChain found 57.3% of its technical respondents already running agents in production. Microsoft has made computer use generally available. MCP and A2A are giving the ecosystem common ways to connect tools and agents.
As seen above, the main constraint is increasingly reliability across a complete workflow. Quality remains the leading deployment barrier in LangChain's survey, while companies are installing observability, permissions, approval gates and agent-specific security controls at the same time that models become more capable.
User behavior is changing too. People can hand coding agents jobs representing hours of human work and supervise several tasks in parallel. Enterprises are beginning to measure agent output through resolved conversations and completed work rather than prompts or seats. Agents are also spreading from engineering into legal, sales, recruiting and other business functions, although those newer categories remain much smaller.
So where are we today? AI agents have crossed the line from an interesting interface into useful software that can execute meaningful chunks of work. They still need boundaries, and unrestricted autonomy remains a poor description of most serious deployments.
The next phase should be much less about proving that an AI can click, code or call a tool. We already have plenty of evidence for that. The real race is to increase the amount of work an agent can finish correctly before a human has to step back in.
| What is changing | Our judgment today |
|---|---|
| Strongest agent market | Coding |
| Fastest large adjacent market | Customer service |
| User behavior | Longer tasks and parallel delegation |
| Enterprise adoption | Production is growing; broad scaling remains limited |
| Software access | APIs first, computer use for the long tail |
| Integration layer | Rapid convergence around MCP |
| Agent-to-agent interoperability | Emerging through A2A |
| Biggest production problem | End-to-end reliability |
| Security direction | Dedicated identities, permissions and audit trails |
| Labor impact so far | Tasks are shifting faster than whole jobs |
| Business model | Moving toward completed work and outcomes |
| Overall verdict | AI agents are becoming a genuine execution layer |
If you want more recent data on this point, please see our latest agentic AI market report.
This chart, included in our agentic AI market deck, shows how autonomous AI agent platform technology has evolved over time
OUR METHODOLOGY
This analysis asks what is actually changing with AI agents now. We broke that question into the dimensions that best show whether the shift is real: capability, reliability, usage, enterprise deployment, commercial traction, integration, security and the way work itself is changing.
We gave the most weight to evidence of real behavior: people delegating longer tasks, agents completing work across several steps, companies putting systems into production, usage expanding after deployment, customers paying for outcomes and technical controls appearing around those deployments. Benchmarks were used mainly to understand capability and reliability rather than as proof of commercial adoption.
Different kinds of evidence were kept in their proper lane. Usage data helped us understand behavior, surveys helped establish how deployment is spreading, revenue and pricing provided evidence of commercial demand, and standards or security work showed what infrastructure is forming around the market. We compared signals when they answered the same underlying question rather than forcing unlike metrics into a single score.
Company-reported data is used where companies have the clearest view of their own revenue, product activity or customer usage, but we treat those figures as company-reported rather than independent market measurements. Broader conclusions rely more heavily on convergence across several companies, independent research, large surveys and standards bodies.
We prioritized recent, checkable sources and original research wherever possible. Key sources include METR on frontier-agent task horizons, Anthropic on Claude Code revenue and usage, OpenAI on longer Codex tasks and parallel delegation, OpenAI on Codex weekly usage, McKinsey on enterprise agent scaling, LangChain on production adoption, quality and observability, Salesforce on customer-service agent adoption, and Salesforce's Agentic Enterprise Index.
Other important sources include Intercom on Fin resolution rates and outcome pricing, Microsoft on computer-using agents in Copilot Studio, the Linux Foundation on MCP and the Agentic AI Foundation, the Linux Foundation on A2A adoption, NIST on agent identity and authorization, OWASP's Agent Control Standard, Anthropic's Economic Index work on automation in Claude Code, OpenAI and Cisco's Codex case study, and The Information on Cognition's latest financing and annualized revenue.
The final judgments come from that aggregation. We put more confidence behind a conclusion when several recent signals from different angles pointed in the same direction, and kept claims narrower where the evidence was narrower. The goal is to separate what agents can demonstrate, what people and companies are already doing with them, and what has genuinely begun to scale.
In our agentic AI market deck, we identify pain points entrepreneurs should prioritize
Related blog posts
- AI Agents: what are the biggest challenges now?
- AI Agents: what are the biggest unsolved problems?
- AI Agents: what is getting real adoption now?
- AI Agents: what is actually working now?
- What are the latest funding developments in the agentic AI market?
