Will persistent AI agents be bigger than coding agents?

Why is this question suddenly serious?

Persistent AI agents have moved from a research idea to a real product race, while coding agents have already become a large software business.

The change is visible across several platforms at once. Microsoft recently made Copilot Cowork generally available for long-running work across Microsoft 365, and Google is rolling out Gemini Spark as a personal agent that can keep working around the clock. Salesforce has pushed Agentforce deeper into its CRM products, while ServiceNow has opened its workflow platform to agents built by other vendors. OpenAI and Anthropic are also taking the execution systems first developed for coding and applying them to documents, research, data analysis and office work.

Several major vendors are making the same move at once, so the comparison is worth asking now. A year ago, coding agents had the clearer product and persistent agents mostly had the broader story. Today, both sides have products, usage and revenue, but coding still has far stronger commercial evidence.

The uncertainty comes from a basic mismatch. Coding agents serve a smaller group of users but solve expensive, structured work. Persistent agents could reach almost every digital worker and consumer, yet they must survive messier data, weaker feedback and much higher trust requirements.

What counts as a persistent AI agent, and what does “bigger” mean?

A persistent AI agent is an AI system that keeps useful state, resumes work later, reacts to new events and can take actions without being prompted through every step.

Remembering a user’s name is too weak to qualify. A persistent sales agent could notice that a prospect replied, update the CRM, prepare the next message and schedule a follow-up. A persistent finance agent could monitor incoming payments, investigate exceptions and reopen an unresolved case the next morning. The useful distinction is continuity across time, tools and changing events.

Coding agents already have some of these properties. GitHub Copilot coding agent works in the background, Claude Code can run long tasks, and Codex can handle several jobs in parallel. We are therefore comparing two overlapping layers: agents centered on software development and agents that persist across broader business or personal workflows.

For this article, “bigger” means more annual revenue and more economically useful work. User counts alone can mislead. A lightly used assistant with 100 million users may create less value than a coding agent used for hours every day by five million developers. We will judge both the current commercial market and the likely mature market.

Which market is bigger today?

Coding agents are clearly bigger today when we look at visible standalone revenue, paid adoption and repeat usage.

Anthropic says Claude Code now generates more than $2.5 billion in run-rate revenue. Cursor has reportedly passed $2 billion in annualized revenue. Those two products alone put visible coding-agent revenue above $4.5 billion before we count GitHub Copilot, Codex, Replit, Devin or the coding revenue hidden inside broader AI subscriptions.

Persistent-agent revenue is growing quickly but remains less proven as a separate category. Salesforce reported $1.2 billion in Agentforce annual recurring revenue in its latest quarter, up from $800 million one quarter earlier. That is a 50% jump in one quarter. Microsoft has more than 20 million paid Microsoft 365 Copilot seats. It also says tens of thousands of companies manage tens of millions of registered agents through Agent 365, while leaving the persistent-execution portion of revenue undisclosed. ServiceNow says the number of Now Assist customers spending more than $1 million annually grew by more than 130%, again without separating agent revenue.

The gap is smaller than a simple $4.5 billion versus $1.2 billion comparison suggests because much of the persistent-agent business is bundled. Even after making that adjustment, coding still leads the current scoreboard.

Measure	Coding agents today	Persistent agents today	Current leader
Visible standalone revenue	More than $4.5 billion from Claude Code and Cursor alone	$1.2 billion for Agentforce, with much more revenue bundled	Coding agents
Paid professional adoption	More than 4.7 million paid GitHub Copilot subscribers, plus other products	More than 20 million paid Microsoft 365 Copilot seats, although most are still broader copilots	Mixed
Daily work intensity	Already used repeatedly inside development workflows	Strong in selected CRM, IT and productivity workflows	Coding agents
Market breadth	Concentrated in software development	Spans sales, service, finance, IT, HR and personal work	Persistent agents

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

Why did coding agents find product-market fit first?

Coding agents reached product-market fit first because software gives AI unusually clear instructions, fast feedback and cheap recovery from mistakes.

A coding agent can read a repository, edit files, run tests, inspect an error and try again. Version control preserves the previous state. A developer reviews the pull request before the change reaches production. Most business workflows offer fewer automatic checks and much less reversibility.

The buyer also understands the value quickly. Software engineers are expensive, and an extra feature, bug fix or migration can affect thousands of customers. A company can justify a $100 or $200 monthly tool when it saves several engineering hours. The same price is harder to defend for a general assistant that occasionally writes an email or reorganizes a calendar.

Coding has another advantage: the work is surprisingly standardized. Companies use different architectures, but developers still share languages, repositories, terminals, tests and deployment systems. Customer service, procurement, finance and recruiting vary much more between organizations. A broad agent must learn each company’s data, rules and exceptions before it becomes useful.

Coding became the first large agent category because its environment gives imperfect models enough structure to create value now. Broader persistent agents still need stronger surrounding systems.

Has coding-agent adoption moved beyond hype?

Coding-agent adoption has moved well beyond experimentation, although intense use remains concentrated among a smaller group of developers.

Microsoft reported more than 4.7 million paid GitHub Copilot subscribers in its previous quarter, up 75% year over year. In its latest quarter, nearly 140,000 organizations were using GitHub Copilot and enterprise subscriptions had almost tripled. Anthropic says Claude Code’s weekly active users doubled from the start of the year, while business subscriptions quadrupled and enterprise customers generated more than half of product revenue.

The behavior inside engineering teams also looks real. A recent study followed tens of thousands of Microsoft engineers. Developers who adopted command-line coding agents merged roughly 24% more pull requests over four months than the model estimated they otherwise would. A merged pull request is an imperfect measure of value, but it is much harder evidence than a satisfaction survey or a benchmark score.

Open-source research adds an important correction. A recent study found agent-generated pull requests across thousands of popular repositories, yet the median repository produced only one or two such pull requests over three months. Heavy use is growing fast and remains concentrated in a minority of repositories.

Coding agents have durable product-market fit today. The remaining debate is how deeply average teams will use them.

Are persistent AI agents already generating serious revenue?

Persistent AI agents are already producing serious enterprise revenue, but most deployments remain narrower than the “digital employee” label suggests.

Salesforce provides the cleanest financial trail. Agentforce annual recurring revenue rose from $800 million to $1.2 billion in one quarter, while agentic work units increased from 2.4 billion to 3.8 billion. More than half of Agentforce and Data 360 bookings came from existing customers, which points to expansion after initial deployment, with new trials playing only part of the role.

Microsoft shows a different kind of scale. More than 20 million paid Microsoft 365 Copilot seats give the company a huge base for persistent work, and Copilot Cowork can now execute long, multi-tool tasks while the user is away. Agent 365 already contains tens of millions of registered agents across tens of thousands of companies. Registration overstates active use, but organizations rarely buy an identity, security and governance layer for a category they expect to disappear.

ServiceNow’s latest results strengthen the pattern. Customers spending more than $1 million annually on Now Assist grew by more than 130% year over year. Google’s Gemini Spark is earlier and more consumer-focused, yet its 24/7 design shows that persistent execution is also moving into mass-market assistants.

The market has crossed the point where we can dismiss persistent agents as demos. Broad autonomous work, however, remains uncommon.

Does the larger potential user base decide the winner?

Persistent agents have a much larger potential audience, but that advantage only matters when broad access turns into frequent paid work.

SlashData estimated 47.2 million developers worldwide. Microsoft already has more than 450 million paid commercial Microsoft 365 seats, roughly ten times that developer population. Add Google Workspace users, frontline workers, small-business owners and consumers, and the possible reach becomes far larger.

Current paid adoption hints at that difference. As seen above, Microsoft 365 Copilot’s paid base is more than four times the previously disclosed GitHub Copilot subscriber base. Many Microsoft 365 users still rely on Copilot for individual prompts, so paid seats overstate persistent-agent adoption.

The coding audience is smaller but much easier to monetize. Developers use agents repeatedly, understand their limits and can verify the output. The broader office population contains many people whose work is less digital, less standardized or less measurable.

User-base size strengthens the long-term case for persistent agents, but it does not hand them the market.

Can persistent agents create more value inside one company?

Persistent agents can eventually create more total value inside a large company because they can touch many departments and keep running between human interactions.

A coding agent mainly expands software capacity. A persistent-agent platform can also qualify leads, update customer records, triage support cases, reconcile transactions, monitor incidents, prepare reports and coordinate recurring work. Each individual workflow may be less valuable than software engineering, but the combined surface is much wider.

Salesforce’s latest figures show how that expansion can work. More than half of Agentforce and Data 360 bookings came from existing customers, while agentic work units rose 111% quarter over quarter. Customers appear to start with one workflow and then consume more work through the same platform. Microsoft is seeing a similar pattern in Copilot Credits, where consumption nearly doubled quarter over quarter as customers added custom agents.

The upside depends heavily on process design. PwC’s latest global AI performance study found that 20% of companies captured 74% of measured AI value. The leading group earned 7.2 times more revenue and efficiency gains than peers. That concentration suggests that installing agents is easy compared with redesigning work around them.

Persistent agents have the higher ceiling per organization. Reaching it will require cleaner data, clearer rules and fewer handoffs than most companies currently have.

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

Which persistent-agent workflows will scale first?

Persistent AI agents will scale first in high-volume workflows where inputs are digital, outcomes are measurable and a human can handle the unusual cases.

Customer-service triage fits well because requests arrive continuously, historical examples exist and unresolved cases can be escalated. IT operations offer similar conditions: agents can monitor alerts, gather diagnostics and apply approved fixes. Sales and recruiting agents can research people, maintain records and trigger follow-ups while leaving final decisions to employees. Salesforce and ServiceNow are already concentrating much of their agent activity in these kinds of structured operational workflows.

Finance operations can also support valuable agents, especially for reconciliation, collections and exception handling. Progress will be slower wherever an agent can move money, sign an agreement or create a regulatory problem. Personal agents face a different obstacle: users must grant access to email, calendars, files, location and sometimes payments before the product becomes truly useful.

Consumer persistent agents may eventually have more users, but enterprise workflows should produce more revenue first. Companies can measure saved labor and faster response times. Consumers often expect similar capabilities inside one general subscription. Google’s early Gemini Spark positioning reflects that consumer opportunity, although the product is still much less commercially proven than enterprise agent platforms.

Workflow	Why it fits persistent agents	Main obstacle	Near-term outlook
Customer service	Continuous demand, clear records, easy escalation	Wrong resolutions can damage trust	Strong
IT operations	Digital signals, repeatable checks, measurable recovery	Security permissions and rare failures	Strong
Sales and recruiting	Recurring research, follow-ups and record updates	Poor targeting and brand risk	Strong
Finance operations	Reconciliation and exception queues are structured	Audit rules and transaction authority	Selective
Personal assistance	Huge audience and many recurring tasks	Privacy, fragmented apps and weak willingness to pay	Promising but slower

Can persistent AI agents work without supervision today?

Persistent AI agents still need supervision for long or consequential workflows, even though their useful task horizon is improving quickly.

METR’s current measurements show that frontier agents can complete longer software tasks than earlier systems. The benchmark is valuable because it measures success against the time a skilled human would need. METR also warns that its tasks are cleaner and easier to verify than much of normal office work. The result transfers poorly into finance, healthcare or customer operations.

Recent professional-work benchmarks expose the gap. DELEGATE-52 found that models introduced silent corruption during long document-editing workflows across 52 domains. A healthcare benchmark released this year found that performance on coding-style long tasks transferred poorly to policy-heavy healthcare operations involving several roles and irreversible decisions. Other desktop benchmarks still describe professional, multi-application work as an open problem.

Length makes small weaknesses compound. An agent with 99% reliability on each dependent step has only about a 37% chance of completing 100 steps without a mistake. Real systems can recover from some errors, but they can also carry one wrong assumption through the rest of the workflow.

Today’s workable model is bounded autonomy. Agents can gather evidence, prepare drafts, update low-risk records and perform approved actions. High-impact steps still need checkpoints.

Can better memory make persistent AI agents dependable?

Better memory will make persistent AI agents more useful, but dependable long-running work also needs verification, permissions and recoverable execution.

A persistent agent needs to remember the current plan, previous interactions, company rules and unresolved events. It also needs to forget or downgrade stale information. Keeping every detail forever can make decisions worse when old prices, preferences or policies remain in the memory store.

The hard part is choosing the right memory at the right moment. A sales agent may remember that a customer rejected one offer six months ago, while missing the newer contract that changed the account. A personal agent may retain a temporary travel preference and apply it to every later trip. Poor retrieval and conflicting records survive even when storage expands.

The latest agent platforms reflect this broader view. Microsoft describes Foundry agents as durable and stateful, while OpenAI’s stateful runtime combines persistence with secure execution and orchestration. Anthropic’s guidance for long-running agents emphasizes progress files, checkpoints and explicit handoffs across context windows.

Memory helps an agent resume. Verification tells the agent whether the work still makes sense. Persistent systems need both.

Will security and governance slow persistent-agent adoption?

Security and governance will slow persistent-agent adoption, and the burden is much heavier than it is for most coding tools.

A persistent agent may read email, customer records, contracts, calendars and financial data while the user is offline. Long-lived access creates more opportunities for a malicious instruction, an outdated permission or a compromised integration to cause damage. The organization must know which agent acted, who approved its authority and how to reverse the result.

Microsoft’s product strategy gives us a useful market clue. Agent 365 is a separate control plane for agent identity, security, observability and compliance, priced at $15 per user per month. A paid governance layer this early tells us that agent sprawl has already become a real enterprise problem.

The constraint will favor incumbents. Microsoft, Salesforce, ServiceNow and Google already control identities, permissions, records and audit trails. A startup may build a smarter agent, but the platform holding the data decides what that agent can actually do.

Governance will delay the market and steer more revenue toward the platforms that already control identity and data.

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

How will companies charge for persistent AI agents?

Persistent AI agents will push software pricing away from a pure seat model toward a mix of subscriptions, usage and business outcomes.

A coding agent still maps reasonably well to a developer seat, although even GitHub is moving toward usage-based AI credits. A persistent agent may work through the night, serve several employees and execute thousands of actions. Charging only for the person who created it would miss much of the value and cost.

Salesforce now offers both employee licenses and consumption pricing. Its Flex Credits price actions, while Agentforce add-ons can provide unmetered usage for licensed employees. Microsoft uses Copilot Credits across Cowork, Copilot Studio, Dynamics agents and Power Platform, with task cost tied to model use, context retrieval, tool calls and runtime. ServiceNow is bundling more AI and governance into its product tiers.

Outcome pricing will appear in narrow workflows such as cases resolved, invoices collected or meetings booked. It will remain difficult when several people and systems contribute to the same result.

The market will look messy in financial reports. Persistent-agent revenue will show up inside productivity suites, CRM, cloud consumption, workflow software, security and professional services. The category may become economically huge while remaining hard to measure as one line item.

Are coding agents turning into persistent general agents?

Coding agents are already expanding into persistent general-work agents, which could let today’s coding leaders capture a large part of the broader market.

OpenAI says Codex now has more than five million weekly active users, up more than sixfold since the desktop application launched. Knowledge workers already represent about 20% of users and are growing more than three times as fast as developers. They use Codex for reports, spreadsheets, presentations, contracts, research and workflow automation.

OpenAI’s separate usage study found that 70.2% of sampled individual users had assigned at least one task estimated above one hour of human work. Another 25.6% had assigned a task estimated above eight hours. Non-developer organizational use grew 189-fold from the previous baseline. The estimates are directional, but the movement away from short coding requests is unmistakable.

Anthropic is following the same path. Claude Code produced more than $2.5 billion in run-rate revenue, then Anthropic used the same agentic foundations for Cowork and role-specific plugins in sales, legal and finance. GitHub’s coding agent already runs asynchronously in the cloud and remembers context within a pull request.

Coding agents may become the gateway to persistent work. They learned planning, tool use, file editing and error recovery inside the environment where those skills were easiest to test.

Could coding remain the biggest specialist agent market?

Coding is likely to remain the biggest specialist agent market even if persistent agents become larger in total.

No other single professional workflow combines such expensive labor, frequent usage, digital inputs and automatic checks. Developers can run several agents at once, use large context windows and consume meaningful compute every day. Revenue per active user should therefore stay high.

Coding demand may also expand as production gets cheaper. When companies can build software faster, they often create more internal tools, integrations and experiments. That generates additional maintenance, testing and review work for agents. The category can keep growing even when fewer human hours are needed for each feature.

Broader persistent work will be split across sales, service, finance, IT, healthcare, legal work and personal assistance. Together, those markets can exceed coding. Individually, most will struggle to match coding’s concentration and spending intensity.

We expect a wide persistent-agent economy with coding as its strongest vertical.

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

What could stop persistent agents from overtaking coding agents?

Persistent agents could lose the race if reliability, data access and organizational change improve more slowly than model capability.

Reliability remains the first barrier. A persistent agent adds cost when it creates exceptions faster than employees can review them. Long workflows, conflicting instructions and changing environments still expose weaknesses that coding benchmarks often hide.

Data and integration come next. Many companies store key context across old databases, spreadsheets, messages and undocumented employee habits. API access leaves those inconsistencies intact. Persistent execution can automate confusion just as efficiently as it automates good process.

Authority creates the third barrier. Companies may happily let agents search, summarize and draft while refusing permission to send, approve, purchase or change records. The largest savings usually appear after the agent receives those powers.

Organizational redesign could be the slowest part. PwC’s current global study found that 20% of companies capture 74% of measured AI value. The gap suggests that a small group has learned how to rebuild workflows, while most organizations still layer AI onto old processes.

A final risk is commoditization. Memory and background execution may become standard features inside general AI subscriptions. Persistent agents could perform enormous amounts of work without producing an equally large standalone software category.

What would prove persistent agents have overtaken coding agents?

Persistent agents will have overtaken coding agents when paid production work, delegated authority and recurring consumption all move ahead of the coding market.

Revenue is the first test. We would want to see several persistent-agent products reach the multibillion-dollar scale already achieved by Claude Code and Cursor, or major platforms disclose enough bundled consumption to show the same result.

Usage is the second test. Registered agents and purchased seats are weak measures unless those agents complete useful workflows repeatedly. Salesforce’s rising work units and Microsoft’s growing Copilot Credit consumption are encouraging, but neither yet gives us a clean industry-wide success rate.

Authority is the decisive test. Persistent agents need to close routine cases, update systems of record, apply approved fixes and complete transactions within limits. Drafting alone is too limited to create the full economic market.

The last test is organizational structure. Companies will budget for agent capacity, assign managers to agent fleets and organize human teams around exceptions. When that behavior becomes normal across several industries, the crossover will be hard to dispute.

Proof point	What we would need to see	Status today
Revenue	Several broad-agent products above $2 billion in annual revenue	One clear billion-dollar product, with more bundled revenue
Repeated production use	Large volumes of completed workflows with disclosed success rates	Growing consumption, limited quality disclosure
Delegated authority	Routine actions completed without per-step approval	Mostly bounded and checkpointed
Organizational redesign	Teams and budgets built around agent capacity	Visible among leaders, uncommon across the market

Will persistent AI agents be bigger than coding agents?

Yes, persistent AI agents will probably become bigger in total, while coding agents stay ahead for now and remain the strongest specialist category.

Coding owns the current market. Claude Code and Cursor alone generate more than $4.5 billion in visible annualized revenue. GitHub Copilot has millions of paying subscribers, and recent evidence shows real changes in engineering output. No broad persistent-agent category can yet match that combination of revenue, daily intensity and verifiable work.

Persistent agents have the stronger long-term arithmetic. Microsoft’s paid commercial productivity base is roughly ten times the estimated global developer population. The same agent infrastructure can spread across customer service, sales, finance, IT, HR and personal administration. Persistent agents can also keep generating usage while no human is sitting in front of the product.

The route will be slower than the market hype suggests. Coding agents operate in a forgiving environment with tests, version control and expert reviewers. Broad persistent agents must navigate fragmented data, vague requests, permissions and decisions that can be hard to reverse. Current benchmarks still show that long professional workflows expose serious weaknesses.

The crossover should happen gradually, workflow by workflow, starting with bounded and measurable operations. Coding agents will keep expanding at the same time, and several of the eventual broad winners may come directly from coding platforms.

Persistent agents should become the larger economic layer. Coding agents will remain the stronger product category for now, the largest specialist market later, and the training ground from which many general agents emerge.

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.