“The process of creative destruction is the essential fact about capitalism.”
— Joseph Schumpeter, Capitalism, Socialism and Democracy (1942)
Coding agents are still underrated. Not by a factor of two. By a factor of ten. The market knows the category exists. It does not yet know that the same category is breaking containment into every other surface of the enterprise, and that the runtime doing the breaking is also the runtime that will power the wave coming after it.
You already know the headline numbers. Cursor went zero to a billion in two years. Claude Code is past two and a half billion in ARR a year after launch. Eighty-four percent of developers use AI tools. None of this is the news.
The news is that what’s visible (the Cursor line, the Claude Code line, the adoption number) is the entry point of a category an order of magnitude larger than what is currently priced in. The data that proves it has been in the public record for a year. Half of Claude’s monetized usage is coding. Sixty percent of Vercel’s admin traffic is non-human. Cursor’s net revenue retention has no SaaS precedent. Each of these, on its own, would be a generational fact. Together they describe a category that’s not adjacent to the rest of enterprise AI. It’s underneath it. The substrate through which AI is reaching everything else.
The argument compounds. Coding agents are the entry point through which AI reaches the firm. Once inside, they form addictive consumption unlike anything in the B2B record. From there they exhibit capability autocatalysis: their scope grows faster than any prior software category. Four more consequences fall out of these three. The service-as-software wave they power. The background-agent fleet you aren’t metering. The capital autocatalysis that funds the next inference call. The decision-agent avatar coming after that. The joint phenomenon has the right name already. Breaking containment. The rest of this essay specifies the mechanism, and the scale.
I. The Entry Point
“A cage went in search of a bird.”
— Franz Kafka, Zürau Aphorisms, 16 (1917)
Coding represents roughly half of Claude’s monetized use cases. Take that seriously. If half of the marginal compute dollars Anthropic moves inside enterprises are denominated in coding, then coding isn’t one AI application among many. It’s the application. The one through which AI is being metered, billed, and embedded into the firm. Claude Code is at $2.5B ARR at the one-year mark. OpenAI’s Codex sits around $2B. There is no other AI product category at that monetized scale, and the gap is widening.
This breaks the standard adoption model. The standard model is top-down. A CIO evaluates platforms, a CISO blesses them, procurement signs MSAs, the workforce eventually uses them. The actual pattern is shadow IT compressed to quarters. Cursor went from zero to $1B ARR in under twenty-four months. No B2B SaaS company has ever done that. Cursor’s customer is the individual engineer with a corporate card, not the procurement officer with a vendor matrix. By the time a CIO writes the policy, the Stack Overflow 2025 figure of 84% of developers using AI tools is already the headcount inside the firm.
The result is a quiet asymmetry. Analysts model which platform the firm will pick. Engineers have already picked, and what they picked reaches everything. CI. Secrets. The data warehouse. The customer-support backlog. The internal docs. The codebase is not an isolated surface. It is the lattice that touches every other surface, because every other surface is, in the end, code. AI did not walk in through the front door. It slipped in through the IDE. Or rather, the IDE went looking for it. Kafka’s cage, not Kafka’s bird.
The historical comparison the analysts reach for is Slack and Dropbox. Bottom-up SaaS that entered the firm on personal cards and outran procurement to billion-dollar revenue. The analogy is correct in shape and wrong in scale. Slack moved between people. Dropbox moved between folders. Coding agents move between systems. An agent installed in the IDE can read the data warehouse, write to the deploy pipeline, restart services, file PRs, post to Slack, query the CRM, because all of those interfaces are, in the firm’s actual structure, code. The codebase is the API to the company, and the agent now holds the key. No prior shadow-IT category had that property. None of them got to call the rest of the stack on the user’s behalf.
II. Addictive Consumption
“The junk merchant does not sell his product to the consumer, he sells the consumer to his product.”
— William S. Burroughs, Naked Lunch (1959)
An addictive consumption pattern has three properties. A dose that is cheap, strong, and immediate. Tolerance, such that older doses feel inadequate. And dependency, such that abstinence costs more than the original benefit. Coding agents are uniquely fitted to enterprise software economics on each of these.
The dose. Cui et al. (2024), a randomized controlled trial of 4,867 developers, found a 26% increase in pull requests completed for an additional $20–$200 per seat per month. That price-to-effect ratio has no precedent in enterprise SaaS. Salesforce, Atlassian, Workday, ServiceNow. None of them, at the margin, has ever offered the buyer that kind of unit economics. When the marginal hit is this cheap and this strong, purchasing stops being a deliberative budget exercise. It becomes habit. Habits do not churn.
The tolerance. Yesterday’s model feels primitive within months. The same task that thrilled an engineer in March feels frustrating in September. The signature is consumption uncorrelated with marginal output. swyx’s framing for the current moment is precise: we are in a phase of capability exploration; people are rewarded for spending more, rather than less. Every model release is consumed not as an upgrade in the usual SaaS sense, but as a refilled prescription. The prescription is what keeps the workflow going.
The dependency. The widely-cited METR result from July 2025, where senior open-source developers ran 19% slower with AI assistance on mature codebases, gets read as evidence that the tooling does not work. That reading misses the mechanism. Engineers who have used coding agents for a quarter cannot revert without their productivity feeling halved, regardless of what a controlled study measures. The substance has formed a workflow dependency, and the cost of withdrawal is a perception cost, not a measurement cost. Perception costs determine renewal behavior. Renewal behavior determines lifetime value. Coding agents have the highest net retention of any enterprise software category, and the discourse has not absorbed why.
III. Capability Autocatalysis
“Technology, collectively, creates itself out of itself.”
— W. Brian Arthur, The Nature of Technology (2009)
The two arguments above describe a tool. This one describes why coding agents are not a tool but a category of their own. Every prior enterprise software category had a fixed scope set by the vendor. CRM stayed CRM. ERP stayed ERP. The product expanded when the vendor shipped a feature; the product did not expand of its own accord. Coding agents are the first software category whose scope is set by what the user can describe, not what the vendor ships.
Ask a coding agent to integrate a new system and it writes the integration. Ask it to reach a new dataset and it writes the connector. Ask it to deliver a tool to a non-engineering team and it writes that tool. The surface area inside the firm grows without a single new vendor signature. Call this the capability autocatalysis of coding agents. It is the property Arthur and Kauffman ascribe to general-purpose technologies, observed inside the enterprise rather than inside a research lab. The lab version gets all the attention. The enterprise version is where the value accrues.
The mundane version is already everywhere if you look. Inside any large firm that has been running coding agents for a year, there is a quiet long tail of agent-built internal tools that no procurement officer signed off on. A Slack bot that triages incidents. An expense-classifier that reads receipts. A meeting-summarizer that posts to the right channel. A script that flags slow CI builds. An internal dashboard that none of the dashboard vendors are billed for. None of these were features in any vendor’s roadmap. Each one displaces a workflow that was previously either manual labor or a paid SaaS line item. Multiply this long tail by the number of engineering orgs running agents in 2026, and you have a substrate of free-form internal automation that no analyst is measuring, because no SKU corresponds to it. The agent is eating not just code but every small workflow the firm forgot to vendor for.
The empirical proof that this is happening at production scale is the line of the section. Vercel reports that roughly 60% of traffic to its admin app is non-human, agent-driven. That is not a research artifact. That is agents calling agents at production traffic ratios. Pair it with the lab track. Anthropic’s Responsible Scaling Policy now formally tracks AI R&D self-improvement as a capability gate. Sakana’s AI Scientist v2 has produced a workshop-accepted ML paper end-to-end. The picture is clear. Capability autocatalysis is closing on two tracks simultaneously. The lab track gets the discourse. The production track gets the trillion dollars.
IV. The Mispricing
“We had the experience but missed the meaning.”
— T. S. Eliot, Four Quartets — The Dry Salvages (1941)
Entry, addiction, and capability autocatalysis are not three independent observations. They are one mechanism observed at three timescales. Quarters. Months. Days. The entry creates the habit. The habit creates the demand for capability autocatalysis. The autocatalysis deepens the entry. The market is pricing each as a separate question and applying SaaS multiples to the result. The correct pricing model is closer to a substrate than to a tool, and substrates capture surplus that tools never can.
The implication runs the other way too. Salesforce, ServiceNow, Workday, the dominant enterprise vendors of the last era, were built around procurement. Their distribution motion is incompatible with the channel coding agents are using. They cannot move at engineer-speed. The substrate will not belong to them.
It will belong to whoever runs the agent loop between the engineer and her tools. Today that means Anthropic and OpenAI, with the second-tier wrappers (Cursor, Cognition, possibly GitHub) in their wake. Within twenty-four months, the question inside every enterprise will not be which AI should we adopt. It will be how do we restructure around an agent workforce we never decided to hire. By 2028, every Fortune 500 firm will have a coding-agent budget larger than its CRM budget. The companies that own the runtime when that happens will not be priced as software companies. They will be priced as the metering layer of the next general-purpose technology.
The bear case is three claims. One: scaling laws bend and model improvements plateau. Two: open-source models commodify the substrate and margins compress. Three: the METR result generalizes, senior engineers become net-negative users of agents, and the consumption pattern collapses. None of the three survives contact with the mechanism. If scaling plateaus, current capability is already at the price-to-effect ratio that has produced the addiction signature; the substrate is locked in even if the model never gets better. If open-source commodifies the model, the runtime does not commodify. Neither does the trust calibration, the audit trail, the embedded workflow, or the security perimeter, and those are what the buyer is actually paying for. If METR generalizes, the perception cost of withdrawal still exceeds the measured benefit of reversion; engineers stay on the substance even when the controlled study says they shouldn’t. The bear case wins the model layer. The bull case wins the substrate.
The data has been in the open for a year. Half of Claude is coding. Sixty percent of Vercel’s admin traffic is non-human. Cursor went from zero to a billion in twenty-four months. Each of these, on its own, would be a generational fact. Together they are the same fact, restated. The analysts have had the experience. They are still missing the meaning.
V. Service-as-Software
“Tout est service contre service.” (Everything is service for service.)
— Frédéric Bastiat, Economic Harmonies (1850)
The bullish version of the thesis lives here. The lazy form of this argument is more developers will use the tools. That part is real, and it is not done. Stack Overflow’s 84% is a count of users, not spend, and per-user spend keeps rising while we are in the capability-exploration phase. Vibe coding is pulling non-engineers (PMs, designers, ops people, founders) into the developer market itself, expanding the denominator the adoption number is computed against. Cursor’s net revenue retention has no historical comp in enterprise software, and the cohort that started a year ago is the cohort with the most reasons to spend more, not less. The developer curve isn’t bending. The headline market keeps compounding.
That is only half of the bull case. The other half, the half the analysts are not modeling at all, is that coding agents are being deployed as the substrate underneath every service-as-software company being built right now.
Service-as-software is the venture industry’s name for the post-SaaS wave. Sierra (customer service), Harvey (legal), Decagon (support), Hippocratic (healthcare), Crescendo and Cresta (sales). Each is selling not a tool that helps a service worker, but the service worker itself, in software form. The bill is not per seat. The bill is per resolved ticket, per closed contract, per billed hour replaced. Foundation Capital pegs the addressable services market at roughly $4.6 trillion globally. An order of magnitude above the global SaaS TAM that coding agents are eating today.
Coding agents are the right form factor for service-as-software, and they are the only form factor that is. Every service-as-software product is mechanically an agent. It takes a request, decomposes it into multi-step actions, calls tools, recovers from errors, returns an outcome. That is the same loop a coding agent runs against a codebase. The only difference is the codebase. Now it is a customer-support knowledge base, a legal contract corpus, a medical protocol, a sales call transcript. The harness Anthropic built for Claude Code is the harness Harvey runs on. The harness Anthropic built is the harness every vertical agent runs on. Cursor’s reasoning loop is the engine under a thousand service-as-software startups that don’t yet exist. The same scaffolding (context management, tool use, error recovery, audit trail) that makes the coding agent work in production is what every service-as-software firm needs, and what no other category supplies. Chatbots cannot do this work. RAG pipelines cannot do this work. Workflow tools cannot do this work. Only the agent loop, originally built for code, can.
This is what makes the 10x conservative. The coding-agent runtime captures revenue twice on every service-as-software firm. Once when the developers use it to build the product (picks-and-shovels). Then again every time the deployed product makes an inference call to deliver the service (substrate). Anthropic, OpenAI, and possibly Google book both lines. The legacy enterprise vendors were never structured to capture either. The services economy is being rebuilt on a runtime that did not exist three years ago, and the runtime belongs to the same five companies that own coding.
VI. Background Agents
“ἁρμονίη ἀφανὴς φανερῆς κρείττων.” (The hidden harmony is better than the visible.)
— Heraclitus, Fragment 54
Most of the coding-agent discourse is about the foreground loop. You ask, the agent answers, you review, you merge. That is the version of agents that fits in a demo. It is not the version that will dominate consumption.
The version that will is the background agent. The agent that runs without a human watching, on a schedule, on a trigger, on a queue. It opens PRs while you sleep. It pages you only when something it can’t resolve has occurred. It refactors legacy code over weeks with no one tracking the diff in real time. It chews through tickets, expense reports, security alerts, log anomalies, vendor renewals. Work that was nominally human work, quietly reclassified as ambient infrastructure.
Background agents are where the consumption curve actually steepens, and people are sleeping on it. Foreground agents are bounded by human attention; a developer can only review so many suggestions per hour. Background agents are bounded by compute budget, which is a much larger number. The same enterprise that pays $200 per seat per month for a foreground coding-agent subscription will, within two years, be paying $20,000 per month for the background-agent fleet doing work the IT helpdesk and the ops team used to do. The seat licence is the loss-leader. The fleet is the business.
The form factor is wrong for most analysts because it doesn’t look like SaaS. It looks like cron. It looks like a workflow that used to be ETL or RPA, now ten times more powerful because the agent can read the unstructured input the legacy automation never could. Whoever owns the agent runtime owns this line too, and most of the consumption growth in coding agents over the next twenty-four months will come from background agents inside companies that haven’t yet bothered to rename them as such.
VII. Capital Autocatalysis
“Money, says the proverb, makes money. When you have got a little, it is often easy to get more.”
— Adam Smith, The Wealth of Nations (1776)
Section III named the capability autocatalysis. There is a twin sitting on the financial side of the same mechanism, and the discourse has not named it. I am going to call it capital autocatalysis, because that is what it is.
SaaS revenue is paid for human consumption of a tool. It is a subscription on a fixed denominator (seats × renewals), bounded by how many humans exist and how many of them want the tool. Coding-agent revenue is paid for work the agent does. A Sierra agent resolves a ticket and bills for the resolution. A Harvey agent answers a legal question and bills for the answer. The bill funds the inference that runs the next answer. Revenue funds compute funds capability funds revenue. The loop runs without a human bottleneck on either side of it.
This is the first software business model that pays for its own R&D out of its own operating revenue in real time, without a markup. SaaS funds R&D out of next quarter’s renewal revenue and a venture round. The agent substrate funds R&D out of this hour’s billed work. The unit of value is no longer the seat. It is the watt-hour of inference allocated to the agent that earns the work. The runtime that converts watt-hours into resolved tickets owns a financial structure no prior software business model has occupied.
Four additional mechanisms ride this same substrate and compound the TAM expansion further. Each, alone, would be a meaningful market line. Together, they make the $5 trillion look like a floor.
The first is the Model Context Protocol. The wire format every existing SaaS company is now retrofitting itself to speak. The vendors are paying their own engineers to make their products more interoperable with their replacement. Every MCP-enabled tool the legacy SaaS world ships is a new line of inference billed back to the runtime, on the runtime’s terms. The 2010s SaaS catalogue is in the middle of a quiet self-disassembly.
The second is the verification market. As agents do more, the bottleneck stops being capability and becomes can you trust what they did. Eval, tracing, audit, observability. These are not a feature category. They are the next ten-billion-dollar middle layer, sitting between the agent runtime and the enterprise that wants to deploy it at scale. The companies that own this (LangSmith, Braintrust, Arize, and the next two we don’t know yet) are derivatives of agent consumption. They go up when the substrate goes up.
The third is default-pick distribution. Resend, launched in 2023, reportedly captures roughly 70% of Claude’s default email-provider recommendations. There is no SEO equivalent for this. There is no paid-acquisition equivalent. It is a new distribution channel. Be in the training corpus before the next cutoff, then get recommended by every agent that runs. It didn’t exist twenty-four months ago. The companies that win this in the next two years will compound for a decade off the work of a small handful of engineers.
The fourth is the end of the demo economy. For all of B2B software history, the demo was here’s how our product could solve your problem. With coding agents, the demo is the working product, built live for the prospect during the call. The legacy vendor cannot match this. Sales cycles compress from months to minutes. What was a six-month procurement is now a thirty-minute build. Here is how our CRM works gets beaten by here is your CRM, built during this meeting, and you can keep it.
None of the four is the headline number for the coding-agent market. Each is multiplicative on top of it. The runtime captures revenue across all of them at once, on the same balance sheet, on the same inference budget. This is what capital autocatalysis looks like in practice, and it is why every conservative model of the coding-agent TAM is wrong.
VIII. The Next Avatars
“Time forks perpetually toward innumerable futures.”
— Jorge Luis Borges, The Garden of Forking Paths (1941)
Coding agents are the first avatar. They will not be the last.
The same containment-breaking mechanism (entry through individual use, addictive consumption, capability autocatalysis) is portable. The next surface it appears on is the decision layer.
A decision agent is defined by what it does that a coding agent does not. A coding agent produces output that a human reviews and merges. A decision agent makes the call itself. It approves the refund, routes the escalation, fires the underperforming vendor, allocates the budget, says yes to the lateral hire. The output is not code. The output is a decision, executed.
The playbook will rhyme. Decision agents will enter through the inbox, the ticket queue, the expense tool. The IDE’s analogues for non-engineers. They will be installed by individual contributors before procurement notices. They will form a habit the manager cannot revert without a productivity collapse the manager cannot afford. And they will self-extend. First approving expense reports, then vendor contracts, then the lateral hires and reorganizations that today are the substance of middle management.
Within five years, the majority of substantive operating decisions inside large enterprises will already have been made by agents by the time a human approver sees them. The human chain of approvals will remain in place. The decision will not be in it. The CEOs who notice this and the CEOs who do not will diverge sharply, and most of them will be the second kind.
This is why owning the coding-agent substrate matters so much. Whoever runs the agent runtime for code today is positioned to run the agent runtime for decisions tomorrow. The customer relationship, the trust calibration, the security perimeter, the audit trail, the billing infrastructure. All of it transfers. The coding-agent companies are not in the coding market. They are in the agent-substrate market, and code is the beachhead the substrate happens to have entered through. The labs already know this. The Cursor SDK is not a developer tool. It is a platform play that turns the IDE into a runtime developers can build agents on top of. Anthropic’s Claude Agent SDK and the Skills format are explicit substrate plays, not coding features. The companies that own the IDE are not pretending they are in the IDE business.
IX. What This Means For You
“The old world is dying, and the new world struggles to be born; now is the time of monsters.”
— Antonio Gramsci, Prison Notebooks (1929–35)
If you are an enterprise buyer, stop running the AI vendor selection as if it were the same kind of decision as Workday versus SAP. The engineers have already chosen. Your job now is not to pick the platform. It is to set the audit boundary around the substance they are already on. Procurement is downstream of the addiction, not upstream of it. The longer you spend pretending otherwise, the more substrate accretes outside your control, and the harder the retroactive policy will be to write.
If you are a founder building on top of agents, you have two layers of demand pulling for you. Coding to ship features, and service-as-software outcome delivery. The mistake is to pick one. Build for both, and the cycle compounds. Every product feature you ship using coding agents is also a service feature you can sell, which funds more coding-agent consumption, which lets you ship faster than the incumbent procurement loop can react. The window in which a new vertical can become the Sierra of its category is open right now and will close inside thirty-six months.
If you are an engineer, the dependency is real, and so is the leverage. The question of the decade is not whether you use the substance. It is whether you stay sharp enough on the underlying craft that, when you need to debug what the agent shipped, you still can. The agents don’t punish abstinence equally across operators. The engineers who win this decade are the ones who use the substance like a power tool, not like a crutch. And who notice the difference.
The containment is breaking. The category is forming. The substrate is being claimed right now, by a small number of companies, in a window that will not stay open. The analysts who were going to tell you when it happened were always going to tell you after.