Are persistent AI agents the next AI gold rush?

What would make persistent AI agents a real gold rush?

Persistent AI agents already qualify as a gold rush because money, talent and platform power are moving in before broad autonomy has proved safe or consistently profitable.

A technology gold rush has three ingredients: a prize large enough to reshape existing industries, a sudden wave of builders and capital, and major uncertainty over who will keep the value. Persistent agents meet all three. Stanford’s latest AI Index found that global corporate AI investment more than doubled during 2025, while generative AI captured nearly half of private AI funding. At the same time, OpenAI, Microsoft, Google, Salesforce and ServiceNow have all released products built around agents that keep working, remember context or react to events without a fresh prompt.

The economic proof is much narrower. McKinsey found that most organizations scaling agents were still doing so in only one or two functions, and scaled use remained below 10% inside every individual function surveyed. We are looking at a market with genuine production use and unusually aggressive investment while the average company is still figuring out where agents can be trusted.

That combination creates a rush. The opportunity is real, the winners are unclear, and plenty of the money entering the category will produce little lasting value.

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

What makes an AI agent persistent?

A persistent AI agent keeps the job alive after the conversation ends.

It stores the state of the work, remembers selected information, waits for events and resumes from the correct point later. A sales agent might monitor new leads every morning, research each company, update the CRM, send an approved sequence, pause for replies and continue the following week. A normal chatbot can help with one step, but it does not own the process across time.

Persistence also requires more than a large context window. Google’s architecture for long-running agents separates workflow state from chat history, saves checkpoints and wakes the agent when an outside event arrives. OpenAI’s workspace agents can use connected apps, retain memory files, run on schedules and continue in the cloud while the user is away. Microsoft’s autonomous agents monitor data and respond to defined triggers in the background.

We should reserve the term for systems that combine continuity with some freedom to decide what happens next. Traditional automation has continuity but follows fixed rules. A one-off research agent may reason through several steps but disappears once the run finishes.

System	Keeps durable state	Chooses between actions	Resumes without a new prompt
Chat assistant	Limited	Sometimes	Usually no
Traditional automation	Yes	Only through preset rules	Yes
One-off AI agent	Usually no	Yes	No
Persistent AI agent	Yes	Yes, inside defined limits	Yes

Why have persistent AI agents become practical only now?

Persistent AI agents became practical when better models met the less glamorous plumbing needed to keep software running for hours, days or weeks.

Model capability is one part of the change. METR’s updated work on frontier agents still shows a steep rise in the length of software tasks they can complete, even after the researchers expanded the benchmark and revised parts of their methodology. OpenAI now reports much stronger results on long-horizon coding and computer-use evaluations, while Anthropic is explicitly designing models and harnesses for work that spans many steps.

The other part is infrastructure. Durable sessions stop progress from vanishing when a server restarts. Event triggers allow an agent to sleep until a customer replies or a document arrives. Sandboxes limit what it can touch. Version histories, evaluations and audit trails make its behavior reviewable. Shared protocols such as MCP and Agent2Agent reduce the amount of custom integration needed to connect models, tools and other agents. The Linux Foundation says Agent2Agent now has support from more than 150 organizations and integrations across the major cloud platforms.

Costs have also become easier to control. A company can route routine monitoring to a smaller model, call a frontier model for the hard decision, cache repeated context and leave the agent dormant during long waiting periods. The agent may own a month-long workflow while consuming tokens during only a few short bursts.

The breakthrough came from stronger reasoning plus ordinary software engineering. Raw model intelligence was never going to make persistence reliable on its own.

Are the big AI platforms already betting on persistent agents?

Yes, every major enterprise AI platform is now building some version of a persistent agent.

OpenAI’s workspace agents run in the cloud, connect to business tools, keep memory files, work on schedules and expose activity to administrators. Its business release notes now describe a central console where companies can inspect connected apps, recent runs, schedules and agent analytics. That is a clear move from personal chat toward managed digital operations.

Microsoft lets Copilot Studio agents respond to events without waiting for a user, operate continuously in the background and use computers when APIs are unavailable. Google has documented durable multi-day agents with persistent sessions and event-driven resumption. Salesforce says more than 18,000 companies currently run Agentforce, and ServiceNow has opened its workflow system so agents built with Claude, Copilot or other tools can take governed actions inside it.

These companies started from very different businesses, yet they are converging on the same components: memory, triggers, permissions, tool access, monitoring and shared workflows. That level of agreement usually means a new software layer is forming.

The convergence also makes the market harder for startups. Persistent agents need access to the systems where work already lives. Microsoft owns email, documents and identity. Salesforce owns customer records. ServiceNow controls large operational workflows. Google owns productivity software and cloud infrastructure. OpenAI has a large user base and a broad model ecosystem. New entrants need more than a pleasant agent interface.

Are companies really using persistent agents, or mostly testing them?

Companies are using persistent agents in production now, although broad deployment remains the exception.

McKinsey’s latest global survey gives the clearest scale check. Among 1,993 respondents across 105 countries, 23% said their organizations were scaling at least one agentic system and another 39% were experimenting. Most companies that had reached the scaling stage were still using agents in only one or two functions. IT and knowledge management led adoption, where tasks such as service-desk triage and recurring research have clear inputs and outputs.

Deloitte’s survey of 3,235 business and technology leaders points in the same direction. Respondents expect much heavier agent use over the next two years, but only 21% currently report mature agent governance. Companies are moving agents into real work faster than they are building the controls needed for large fleets.

Vendor figures confirm meaningful activity but require caution. Salesforce says more than 18,000 companies run Agentforce. OpenAI says workspace agents can already be scheduled, shared and managed through enterprise controls. These are product-adoption figures rather than independent proof of ROI, and an enabled agent may still handle very little work.

The honest picture sits between the hype and the backlash. Production use is common enough to establish a market, while enterprise-wide autonomy is still rare.

Adoption stage	Best current evidence	What it tells us
Experimenting	39% in McKinsey’s survey	Interest is already mainstream
Scaling somewhere	23% in McKinsey’s survey	Production use is substantial
Scaling inside one function	Below 10% in every function surveyed	Deployment remains narrow
Mature agent governance	21% in Deloitte’s survey	Controls lag behind adoption

Where are persistent agents already earning their keep?

Persistent agents are already earning their keep in customer support, IT operations, sales follow-up and other workflows where success can be counted.

Customer service currently has the strongest public evidence. Microsoft reports that a network of specialized agents on its commercial website cut response latency by as much as 61% and reduced human escalations by as much as 70%. Salesforce says its customers resolve roughly half of inquiries autonomously, while Fisher & Paykel reports that agentic self-service doubled its resolution rate. ServiceNow says Bell improved customer response time by 25% after agents began checking case completeness, filling fields and catching duplicates. These figures come from vendors and their customers, so they are strong case studies rather than market averages.

A fresher example comes from Cars24. OpenAI says the vehicle marketplace now handles more than one million monthly conversation minutes through voice and chat agents and recovers 12% of leads that would otherwise have been lost. The useful part is the loop: the agent keeps context across calls, follows up and ties the work to recovered demand instead of merely answering questions.

These examples share the same shape. The work happens frequently, the agent has access to authoritative records, the next action belongs to a limited set, and a human can take over when confidence falls. A company can measure resolution, escalation, response time, recovered revenue or cost per case.

The weak early use cases usually lack those properties. An agent asked to “help with strategy” has no clean finish line and no easy way to prove it improved the decision. Agents become much easier to buy when both buyer and vendor can agree on what a completed job looks like.

Why did coding agents break through before other persistent agents?

Coding agents broke through first because software gives them constant feedback on whether the work actually functions.

Code can compile, tests can fail, a browser can reveal a broken interface and version control can reverse a bad change. Most office work has weaker feedback. A polished market report can contain a false assumption without triggering an error message. A poor customer email may look acceptable until it damages the relationship.

Recent usage data shows how quickly coding agents are moving beyond autocomplete. A study of Codex activity found that active users grew more than fivefold in the first half of 2026. More than 10% of users managed at least three concurrent agents during a typical week, 26.6% used reusable skills, and the share assigning at least one task estimated to take an experienced person more than eight hours rose almost tenfold. The growth also spread beyond software developers into legal, research and other professional roles.

Anthropic’s long-running engineering work explains why the environment matters so much. Its agents perform better when the harness leaves tests, progress files, Git history and explicit requirements for the next run. In one experiment, parallel Claude agents completed a 100,000-line C compiler over nearly 2,000 sessions, with tests continually forcing the system back toward a working result. That experiment cost about $20,000 in API usage and still required careful harness design, but it proves that persistent agent teams can produce a large, verifiable artifact.

Coding gives us the first convincing commercial model for persistent work. Other industries will follow fastest where they can build similar feedback loops around reconciled accounts, resolved claims, booked appointments, validated compliance checks or completed transactions.

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

Are agents getting good enough for work that lasts several days?

Agents are improving fast enough to take on longer work, but several days of elapsed time still says little about how many decisions they can make safely.

An agent can manage a two-week onboarding process by waking only six times. That is much easier than writing code continuously for 20 hours or making hundreds of linked decisions without review. We should measure the active reasoning burden, the number of irreversible actions and the quality of the feedback rather than the calendar duration alone.

METR’s updated time-horizon research still shows exponential progress on a benchmark of more than 200 software tasks. The organization also warns that the exact estimates are sensitive to task design and modeling choices. One correction reduced the estimated horizon of recent models by as much as 20%, even though the broad upward trend survived.

Frontier-model results are also rising sharply. OpenAI reports that its latest model reaches 82.7% on Terminal-Bench 2.0 and 58.6% on SWE-Bench Pro while using fewer tokens than its predecessor. Anthropic’s latest model is marketed around stronger long-running agent performance, with early enterprise testers reporting fewer turns and tool calls on difficult analytical work. Company benchmarks deserve independent replication, but the direction across model providers, agent products and user behavior is consistent.

Agents will handle longer jobs every year. Benchmark gains, though, will not translate neatly into reliable autonomy inside messy businesses.

Does long-term memory make persistent agents smarter or more dangerous?

Long-term memory makes persistent agents more useful and gives their mistakes a much longer life.

The key risk appears when an agent writes something into durable state. A recent 1,600-task benchmark tested whether personal agents would accept a user’s questionable claim, store it and reuse it later after the original conversation had disappeared. Downstream failure rose from 45.0% in session-only cases to 71.9% after the claim had been committed to memory, an increase of 27 percentage points. The stored claims often lost their original attribution or became broader than what the user had actually said.

That result explains why “more memory” is a poor product goal. A useful system needs to decide what may be stored, where it came from, how certain it is, when it expires and whether a human can correct it. Google’s long-running-agent guidance similarly separates durable workflow state from raw conversation history because replaying everything creates noise, cost and misleading context.

Well-structured memory can still be a major advantage. A research case study covering 96 active days recorded 502 memory-related files and found that 82.9% of measured token volume came from cache reads. The case involved one investigator, so it cannot prove broad productivity gains. It does show how a persistent environment can reuse prior work cheaply when the stored information has clear roles and governance.

The valuable asset will be trusted state: verified facts, workflow checkpoints, permissions, corrections and records of what worked. Companies that preserve every interaction will accumulate a large and increasingly unreliable memory dump.

Can companies trust persistent agents to act without approval?

Companies cannot currently trust persistent agents with broad authority, although many narrow actions are already safe enough to automate.

A 2026 reliability study evaluated 14 agentic models across consistency, robustness, predictability and safety. The researchers found that recent capability gains produced only small improvements in overall reliability. An agent can score higher on the main task while still behaving inconsistently across repeated runs or failing badly after a small change in conditions.

Corporate confidence reflects that problem. Deloitte found mature agentic governance in only 21% of surveyed organizations. In a more recent survey of 200 North American CFOs at companies with at least $1 billion in revenue, 43% felt confident in their AI governance while 53.5% felt only somewhat confident. Cost uncertainty was the most common internal concern, followed by weak confidence in using AI for important operations.

The rule is straightforward: the bigger the consequence, the less freedom the agent gets. Reading a public website, drafting a summary or opening a low-priority support ticket can happen automatically. Sending money, deleting records, changing a contract or rejecting a customer needs tighter limits and often human approval. Anthropic’s recent containment work follows the same logic: cap the agent’s blast radius, isolate environments and make sensitive actions harder to reach.

Trust will grow through boring controls rather than a sudden moment when models become perfectly reliable. Permissions, action limits, tests, audit trails, reversibility and clear escalation rules will decide how much autonomy companies allow.

Can always-on agents become cheap enough for normal businesses?

Always-on agents can be economical because they spend much of their time waiting, but poorly designed agents can create bills that are hard to predict.

Google’s long-running architecture lets an agent save its state, scale down and wake when an event arrives. A procurement workflow may last three weeks while using model inference only when a vendor responds, an approval changes or a deadline arrives. Caching and model routing reduce the cost further.

The market is still experimenting with the right charging unit. OpenAI has moved workspace-agent usage toward token-based credits. Microsoft warns that event triggers affect billing. Salesforce has gone in another direction with a help agent that charges only when it resolves the issue from start to finish; an escalation or negative customer response produces no resolution charge. That pricing forces the vendor to share some of the performance risk.

CFOs are watching this closely. Deloitte’s latest CFO survey found that 46% named cost uncertainty or poor transparency as their largest internal concern about AI. Consumption can jump when an agent retries, loops, calls several tools or escalates every difficult task to an expensive model.

The number buyers will care about is cost per completed outcome. Token prices help developers, but a support director cares about the cost of a resolved case and a finance team cares about the cost of a reconciled account. Agents become attractive when the full outcome costs less after review, failures and integration are included.

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

Will persistent agents replace workers, software seats, or both?

Persistent agents will first reduce the amount of human labor needed for routine workflows, then weaken software pricing based purely on employee seats.

McKinsey found that a median 17% of respondents had already seen AI-related workforce declines inside business functions, while a median 30% expected declines over the following year. Across whole companies, 32% expected an overall reduction of at least 3%, 43% expected little change and 13% expected an increase of at least 3%. Those figures cover AI broadly, so they cannot be assigned entirely to persistent agents. They still show that companies increasingly expect AI to change staffing levels rather than merely assist every existing role.

The first effect will usually appear in hiring and workload, not mass replacement. A support team may handle twice as many conversations without doubling headcount. An accounting team may close the books with fewer temporary workers. Engineers may supervise several coding agents and spend more time reviewing architecture. People remain involved, but fewer hours are needed for each completed process.

Software pricing faces a parallel shift. Because a machine agent can call APIs, move across several products and execute hundreds of small actions, it often needs neither the same interface nor the same license as a human employee. Salesforce’s move toward pay-per-resolution gives a glimpse of where the market may go: vendors charge for work completed, consumption or business outcomes alongside traditional seats.

Seat-based software will survive wherever humans still spend substantial time inside the product. The pressure will be strongest in back-office tools that agents can operate headlessly with little human interaction.

Why are investors throwing so much money at persistent agents?

Investors are throwing money at persistent agents because they see a chance to tap labor budgets as well as software spending.

Three major rounds show where the conviction is concentrated. Sierra raised $950 million at a valuation above $15 billion for customer-service agents. Parloa raised $350 million at a $3 billion valuation for voice-based customer service. LangChain raised $125 million at a $1.25 billion valuation for agent development, evaluation and observability. Together, those rounds total $1.425 billion. Sierra and Parloa captured just over 91% of that amount, which tells us investors currently prefer agents tied to a large operating budget over another broad horizontal builder.

The size of the prize explains the valuations. Customer service, software development, finance operations, sales and compliance consume vast payroll and outsourcing budgets. A vendor that charges for completed work can expand revenue as the customer delegates more volume, even when employee headcount stays flat.

There is also a platform argument. Once an agent becomes the layer through which a company handles customers or runs a workflow, it can gather feedback, improve its process and become expensive to replace. Investors are betting that a few early leaders will turn those operational positions into durable distribution.

The risk is obvious. Valuations assume that pilot activity will convert into recurring, high-volume work before incumbents bundle similar capabilities into existing contracts. Plenty of well-funded agent companies will discover that impressive autonomy is easier to demonstrate than to sell at healthy margins.

Company	Main agent market	Recent round	Reported valuation
Sierra	Customer service and customer operations	$950 million	Above $15 billion
Parloa	Voice customer service	$350 million	$3 billion
LangChain	Agent building, evaluation and observability	$125 million	$1.25 billion

Can agent startups really beat Microsoft, Google and Salesforce?

Agent startups can beat the giants inside specific workflows, while the broad horizontal platform race strongly favors incumbents.

Microsoft, Google, Salesforce and ServiceNow already control the identities, records, approvals and interfaces an enterprise agent needs. OpenAI brings a large user base, frontier models and a growing set of connected business tools. When a generic research, support or productivity agent becomes good enough, these companies can include it in products customers already use.

Startups have a better opening when they own a painful outcome across several systems. Sierra can focus every part of its product on resolving customer issues. Parloa can optimize real-time voice interactions, simulation and call-center deployment. A claims agent, clinical documentation agent or tax agent can carry industry-specific rules and evaluation methods that a horizontal platform will struggle to match quickly.

Open standards reduce one barrier and remove another moat. ServiceNow now allows outside agents to take governed actions through its platform, while Agent2Agent and MCP make cross-platform connections easier. A startup gains access to enterprise systems without rebuilding every connector, but competitors gain the same advantage.

The strongest independent companies will own a workflow, a proprietary data loop, a regulatory advantage or trusted distribution. A thin layer of prompts, memory and connectors will be copied or bundled away.

Where will the easiest money in persistent agents be made?

The easiest money will probably be made by companies that help enterprises control agents, because every serious deployment needs that layer regardless of which model wins.

A company running hundreds of agents needs to answer basic questions quickly. Which agent touched this customer record? Which instruction led to the action? How much did the run cost? Did performance fall after a model update? Can the action be reversed? Ordinary application monitoring does not provide enough detail when software is choosing its own sequence of steps.

LangChain’s growth shows the demand. The company says LangChain and LangGraph reached a combined 90 million monthly downloads, 35% of Fortune 500 companies used its services, and trace volume on the commercial LangSmith platform grew twelvefold in one year. Those are company-reported figures, but they point to a large developer need around evaluation, tracing and deployment.

Governance products should also grow quickly. Deloitte found that roughly four out of five organizations lacked mature agent controls, even as most respondents expected agent use to expand. Identity, least-privilege access, memory approval, cost limits, anomaly detection and audit trails will become standard purchases once agents handle sensitive work.

Payments and interoperability form a third opening. Agents will need secure identities, scoped credentials and ways to transact without receiving permanent access to every account. The companies that make machine work observable, payable and reversible may capture steadier value than the companies selling the most charismatic digital employee.

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

What could bring the persistent-agent gold rush to a halt?

A major gap between impressive demos and dependable economics could cool the persistent-agent market very quickly.

Reliability is the biggest technical threat. Longer workflows create more chances for an agent to misunderstand the goal, store a bad assumption, choose the wrong tool or continue after it should stop. Current research finds that capability is improving faster than consistency and safety. A handful of costly failures could make regulated companies freeze deployments.

Weak returns are the commercial threat. McKinsey found that only 39% of organizations using AI reported enterprise-level EBIT impact, while the strongest performers were much more likely to redesign workflows rather than place AI on top of old processes. Persistent agents will disappoint when companies automate a broken process and keep every existing approval, handoff and piece of software around it.

Bundling could wipe out many startups even if the technology succeeds. Microsoft, Google, Salesforce, ServiceNow and OpenAI can absorb generic agent functions into broader platforms. Customers may welcome the capability while refusing to pay a separate vendor for it.

Costs could also rise faster than expected. Agents often generate many more model calls than a normal chat, and their retries, tool usage, monitoring and human review all add up. The latest CFO data shows that cost uncertainty is already a leading concern.

None of these problems will erase persistent agents. They can easily shrink the number of companies that make money from them.

Are persistent AI agents the next AI gold rush?

Yes, persistent AI agents are the next AI gold rush, although most of the lasting wealth will come from narrow, controlled workflows rather than free-roaming digital employees.

The rush has already started. Major platforms now offer memory, schedules, triggers, cloud execution and agent governance. Nearly two-thirds of organizations in McKinsey’s survey were either experimenting with agents or scaling them somewhere. Customer-service and coding agents are producing measurable work. Investors have placed billion-dollar bets on category leaders before the market’s economics have settled.

The current limits are just as clear. Scaled adoption remains narrow, durable memory can preserve bad information, and better benchmark scores have not removed inconsistent behavior. Companies can delegate bounded tasks now. Giving one agent broad authority across a business would still be reckless.

Three groups should take most of the value. Large platforms will own the general runtime and access to business systems. Vertical agent companies will win where they can prove a completed outcome in one expensive workflow. Infrastructure providers will sell evaluation, security, identity, observability and cost control to everyone else.

Generic “AI employee” startups face the worst odds. Memory, connectors and model access are becoming standard. A durable company will need to own the work, the data or the trust around a specific process.

Our judgment is direct: persistent agents will create a very large software and automation market, and the investment frenzy around them is justified. The fantasy of a dependable autonomous colleague who can handle almost anything remains premature. The money today is in agents that know exactly what job they own, what they may touch and when they must call a human.

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.