What’s worth building in AI coding?

Is there actually still room for new AI coding startups?

Yes. AI coding still has a lot of room, although the easy part of the market has become much harder to enter.

The category is much bigger now than it was even a year ago. JetBrains' latest Developer Ecosystem Survey found that 90% of professional developers were using AI coding agents at work at least weekly and 68% were using them every day. Meanwhile, some of the companies that captured the first wave have become enormous. Anthropic says Claude Code is above $2.5 billion in run-rate revenue. Cursor is now part of SpaceX after a $60 billion acquisition. Cognition, the company behind Devin, is reportedly approaching $1 billion in annualized revenue.

Those numbers could make AI coding look finished. We think they show something slightly different. Developers can suddenly produce far more software than before, and the rest of the development process has not caught up. Code still needs to be reviewed, tested, secured, migrated, deployed, understood and maintained.

The strongest openings today sit around those new bottlenecks. Verification is one. Legacy modernization is another. Secure agent execution, private coding environments, AI cost control and narrowly specialized engineering agents also look much less settled than the assistant market.

So there is plenty left to build in AI coding. We would just start somewhere other than the text box where developers ask an AI to write code.

Can another Claude Code or Cursor still break through?

Probably, but we would not choose that fight without an exceptional technical advantage. General AI coding agents are currently one of the hardest parts of AI coding for a new startup to enter.

JetBrains' latest survey shows just how quickly the leaders are pulling ahead. Claude Code was being used at work by 39% of professional developers surveyed, up from 18% earlier in the year. Codex went from 3% to 16%. GitHub Copilot stood at 21%, while Cursor was at 12%.

Commercial scale is even more intimidating. Anthropic says Claude Code has passed $2.5 billion in run-rate revenue. Microsoft disclosed 4.7 million paid GitHub Copilot subscribers earlier this year, up 75% year over year. Cursor had already passed a reported $2 billion annualized revenue run rate before SpaceX agreed to buy the company for $60 billion. Cognition disclosed a $492 million annualized revenue run rate in May, and Bloomberg later reported that it was already approaching $1 billion.

A new horizontal coding agent therefore needs much more than a good interface and access to strong models. The company needs to solve a class of software work dramatically better than Claude Code, Codex, Copilot and Devin, or find distribution those products cannot easily reach.

We would still take the bet seriously if a startup had a genuinely new technical approach. Another terminal agent, IDE fork or model wrapper whose main pitch is nicer UX is a much harder sell.

AI coding product	Recent scale	What makes entry difficult
Claude Code	$2.5B+ run-rate revenue; 39% workplace usage in JetBrains' survey	Rapid developer adoption plus Anthropic's model advantage
GitHub Copilot	4.7M paid subscribers	GitHub distribution and deep enterprise integration
Cursor	Acquired by SpaceX for $60B	Established developer brand plus large-scale compute access
Cognition / Devin	Annualized revenue reportedly approaching $1B	Strong enterprise position in long-running engineering tasks

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

Is writing code itself becoming a commodity?

Increasingly, yes. Good AI-generated code is becoming easier to access, so startups need a stronger reason to exist than simply generating it.

One clue is how quickly coding products have become multi-model. GitHub's Agent HQ now lets developers run Copilot, Claude and Codex inside the same GitHub workflow. Factory sells enterprise coding agents while switching between foundation models. Niteshift recently launched around the idea that companies should be able to change models without rebuilding their AI engineering infrastructure.

Open models are narrowing the gap as well. Mistral's Devstral 2 scored 72.2% on SWE-bench Verified when it launched, while the smaller 24-billion-parameter Devstral Small 2 can run locally. Mistral's own coding CLI can be pointed at privately deployed models without sending code outside the customer's environment.

Model quality still matters enormously, of course. A better model can turn a frustrating agent into a useful one overnight. The problem for a startup is that another company can often gain access to that same model.

We therefore prefer AI coding businesses where model improvement helps the company rather than erases its differentiation. A migration system with years of proprietary migration data becomes better when Claude improves. A testing platform can run stronger agents when Codex improves. A security layer can inspect output from every model.

Owning the generator alone is getting harder to defend. Owning the workflow around the generator looks much more durable.

Has verification become the real bottleneck in AI coding?

Yes. Right now, AI can create software changes faster than most engineering organizations can confidently approve them.

Google's DORA research captures the tension unusually well. Its 2025 study found that AI adoption had started to improve software-delivery throughput, reversing an earlier negative relationship, while delivery stability was still getting worse as AI use increased. DORA's follow-up work describes developers saving time on creation and then spending some of that time auditing and verifying what the AI produced.

Stack Overflow found the same trust problem from another angle. In its 2026 pulse survey of 1,100 developers and working professionals, 59% were already using agents at work, almost double the previous year's 31%. Yet 63% said they rarely or never let agents operate completely on autopilot.

That gap is one of the strongest startup opportunities in AI coding. Companies increasingly know how to generate a change. They still struggle to answer whether the change matches the original request, breaks another service, introduces a subtle regression or is safe enough to deploy.

If an engineering team goes from producing 100 meaningful changes a week to 500, review capacity can become the constraint almost immediately.

The next big developer platform may spend surprisingly little time helping engineers type code. It may spend most of its time proving that AI-generated changes deserve to ship.

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

Is AI code review already too crowded?

Basic AI code review is crowded now, while deeper software verification still looks very open.

CodeRabbit shows how quickly the category has grown. The company says it now serves more than 17,000 customers, conducts more than two million reviews each week and grew revenue more than fivefold over the past year. It recently raised $143 million at a $1.5 billion valuation.

There are plenty of competitors. Qodo raised $70 million this year around code review, testing and governance. Greptile has built an independent review layer around codebase understanding and reported reviewing more than 500 million lines of code in a single month. GitHub, Cursor and the major coding agents are also building review directly into their products.

We would avoid another reviewer whose main output is a list of comments under a pull request. That feature is getting easier to bundle.

The interesting part starts when the system understands the intended change, checks behavior across several services, knows the company's architectural rules, runs the software and decides how much evidence a human needs before approving it. There is also room around regulated environments where companies need an auditable record of why a change was accepted.

Code review is already becoming a market. Software verification is a much larger problem.

Layer	How open does it look?	Why
Generic PR comments	Crowded	CodeRabbit, Qodo, Greptile and platform vendors already compete here
Cross-repository validation	Open	Large systems create failures that diff-level review misses
Requirement-to-code verification	Very open	Few tools can prove that the implementation matches the original intent
Organization-specific engineering rules	Open	Companies have standards and historical decisions generic models do not know
Regulated change approval	Very open	Auditability and evidence matter as much as finding bugs

Can AI testing become a huge standalone market?

AI testing looks like one of the strongest standalone opportunities in AI coding today because executing software gives us something model-generated opinions cannot: evidence.

Blacksmith is a useful recent example. The company grew from more than 700 customers to more than 5,000 in less than a year. Its valuation jumped from about $60 million to $550 million over the same period, and management says revenue went from roughly $3.5 million in annual recurring revenue last year to tens of millions today. Some customers are already spending more than $1 million a year.

Blacksmith started by making continuous-integration workloads faster. It has since expanded into Codesmith, an agent that can diagnose and repair failed checks. That progression makes sense. More agent-generated changes mean more builds, more tests and more failures to investigate.

The deeper opportunity goes well beyond generating unit tests. An AI testing system could create a realistic environment, work out which behavior needs to be checked, generate missing tests, execute them, inspect failures, reproduce production conditions and send the coding agent back to fix whatever went wrong.

That loop is much closer to autonomous engineering than another code generator. An agent that writes software but cannot reliably tell whether the software works still needs heavy supervision.

A testing platform that closes that loop could become one of the core infrastructure layers underneath agentic software development.

Is security for AI coding agents still wide open?

Yes. AI coding security is still open, especially around what agents are allowed to access and do while they work.

Some basic security functions are already getting bundled. GitHub now automatically runs CodeQL, dependency checks and secret scanning against code created by third-party coding agents such as Claude and Codex. GitHub says those protections had already prevented hundreds of potential vulnerabilities and secret leaks when they were used with its own Copilot coding agent.

That raises the floor for any startup trying to sell ordinary vulnerability scanning for AI-generated code.

The harder security problem happens while the agent is working. Coding agents may read private repositories, execute shell commands, install packages, access databases, use credentials, call internal tools and sometimes interact with infrastructure. Stack Overflow's recent agent survey found that 60% of respondents prevent agents from making unapproved system changes. Developers clearly want more autonomy, but they still worry about what happens when an agent receives too much access.

We see room for products that give agents temporary credentials, sandbox risky execution, enforce least-privilege access, restrict tools, record actions and require approval only when an agent crosses a meaningful boundary.

That could become the identity-and-access-management layer for software agents. If thousands of autonomous engineering tasks are eventually running inside a large company at the same time, controlling what each agent can touch becomes a serious infrastructure problem.

Is legacy software modernization the clearest enterprise opportunity?

Legacy modernization is one of our favorite AI coding wedges right now because companies already have the work, already understand the cost and can usually tell when the job is finished.

Cognition's growth gives us a useful clue. The company disclosed a $492 million annualized revenue run rate in May, with enterprise usage of Devin growing 50% month over month for six consecutive months. More recently, Bloomberg reported that Cognition's annualized revenue was approaching $1 billion.

Scott Wu has also been unusually specific about what enterprises give Devin to do. A lot of the work involves bringing old software up to date, migrating applications between platforms and handling long-tail engineering tasks that human developers tend to postpone.

That is a much better startup wedge than vaguely promising to make every developer faster. Large companies have years of Java upgrades, database migrations, framework transitions, dependency remediation, cloud moves and internal-platform changes sitting in backlogs. The projects are repetitive enough for agents to help, yet contextual enough that ordinary scripts often fail.

We would go even narrower at the beginning. A startup might become exceptionally good at upgrading large Java estates, migrating Oracle workloads, modernizing .NET applications, cleaning up SAP customizations or moving one specific database technology to another.

Each completed project produces useful failure data. Over time, the company learns which transformations break, which tests matter and which edge cases general coding agents repeatedly miss.

That kind of specialist could become much harder to replace than a generic coding assistant.

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

Can private AI coding become a real market of its own?

Yes. Private AI coding looks increasingly credible now that capable coding models can run inside infrastructure controlled by the customer.

Mistral's Devstral Small 2 is a good example. The 24-billion-parameter model is designed for agentic coding and can be deployed locally. Mistral's documentation explicitly supports fully offline setups for organizations that want to keep code and inference inside their own environment.

Startups are beginning to build around that need. Niteshift, founded by former Datadog engineers, recently raised $7 million for an AI coding cloud designed to route work between frontier and open models rather than locking customers into one provider. The founders are betting that companies will want the coding-agent equivalent of a multi-cloud architecture.

Enterprise vibe-coding products are moving the same way. Superblocks recently signed a multiyear collaboration with AWS that lets business users create AI-generated applications inside the customer's own AWS environment, with databases, model access, auditing and security remaining under corporate control.

The opportunity is broader than hosting an open model behind a firewall. AWS, Microsoft and other infrastructure providers can do that themselves.

A stronger private-coding product would manage the whole environment: models, sandboxes, internal packages, permissions, secrets, repositories, evaluation, audit logs and deployment rules.

For banks, governments, defense companies and other organizations with sensitive software, that could become their internal operating system for AI engineering.

Is managing AI coding spend becoming a startup opportunity?

Surprisingly, yes. AI coding costs are becoming large and messy enough that companies are starting to need a financial control layer.

JetBrains recently published its own experience managing this problem. The company said its AI spending increased roughly tenfold during the first half of 2026 as developers adopted tools such as Claude Code and started consuming much more inference. Usage began rising so quickly that JetBrains moved to enterprise arrangements and had to rethink how employees received access to different coding tools.

The underlying problem is easy to recognize. A developer may use Claude Code for one task, Codex for another, a local model for something cheaper and several background agents at once. Pricing also shifts from simple monthly seats toward tokens, premium requests, agent sessions and compute.

A company with thousands of engineers can quickly lose track of which work is expensive, which models are delivering value and which teams are burning inference without producing more shipped software.

We would not build another dashboard that merely reports token spending. The more useful product would route tasks based on cost and difficulty, set team budgets, choose cheaper models when they are good enough and connect AI spend to engineering outcomes.

That becomes especially interesting if model prices keep moving quickly. Enterprises may eventually manage AI coding the way they manage cloud infrastructure today, with FinOps-style controls sitting between developers and several providers.

Is multi-agent coding orchestration still an open market?

Generic multi-agent orchestration already looks vulnerable to bundling, so we would only build it around a much more specific workflow.

Stack Overflow's latest data suggests the market is earlier than the hype makes it sound. Among respondents already using agents, 69% were still working with a single agent. Only 17% used multiple specialized agents and 16% used overlapping or coordinated agents. Sixty-eight percent preferred predictable single-agent setups over more complicated multi-agent configurations.

At the same time, GitHub has already made multi-agent orchestration part of Agent HQ. Developers can assign work to Copilot, Claude and Codex from GitHub, let those agents work asynchronously and review everything through the same repository workflow.

JetBrains is heading in a similar direction with its own agent infrastructure. Cursor has also pushed toward background agents and automated tasks.

A startup whose whole product is a screen where developers launch several agents could get squeezed quickly.

The coordination problem itself is real. Large engineering organizations will eventually need to manage dependencies between hundreds of agent tasks, prevent two agents from changing the same system at once, allocate compute, stop runaway jobs and know which task should be retried.

We just think that capability works better inside something customers already need. A migration platform can coordinate migration agents. A testing company can coordinate repair agents. A security platform can coordinate remediation agents.

Multi-agent orchestration looks more attractive as an engine than as the entire company.

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

Is there still room in vibe coding?

There is room, but horizontal prompt-to-app creation has become one of the busiest parts of AI coding. We would now look downstream, where those generated applications need to become real software.

Lovable recently crossed $500 million in annualized revenue and says people are creating around one million new projects each week. The company has since raised another $400 million at a $13.3 billion valuation and says it now hosts 60 million projects. Base44 reached roughly $150 million in annual recurring revenue only months after crossing $100 million.

Those numbers show that creating software through natural language has become a real behavior, including among people who would never call themselves developers.

They also make another generic app builder difficult to justify. Lovable, Replit, Base44 and several large AI platforms already compete aggressively for the moment when somebody types, “Build me an app that does this.”

The messy part comes later. Somebody needs to decide who can access the application, where its data lives, whether its dependencies are safe, how it connects to internal systems, what happens when usage grows and who maintains it six months later.

Superblocks' recent AWS partnership points toward where this market could go. Its pitch is effectively enterprise vibe coding inside the customer's existing cloud, with IT retaining control over data, security and deployment.

We would rather build around that production layer than another beautiful prompt box. There is likely much more money in helping 10,000 AI-generated internal applications survive inside a company than in helping employees create their first demo.

Are vertical AI coding agents more interesting than general ones?

Yes. Vertical coding agents are one of the strongest ways a new startup can still break into AI coding today.

A vertical agent can own a job rather than compete for the developer's general attention. An agent might specialize in dependency upgrades, CI failures, security remediation, data-pipeline maintenance, ERP customization, cloud migrations or one particular type of enterprise application.

That changes how the product gets measured. Customers do not need to decide whether the AI “feels helpful.” They can measure how many vulnerabilities disappeared, how many applications were migrated, how many failed builds were fixed or how much engineering backlog was removed.

Several companies are already drifting in this direction. Cognition says Devin is heavily used for migrations and maintenance. Blacksmith has moved from running CI workloads into automatically diagnosing and fixing failed checks. Qodo is expanding code review into testing and governance.

Specialization also gives the startup somewhere to accumulate proprietary knowledge. After thousands of migrations, the company knows which transformations tend to fail. After millions of test runs, it knows what normally causes a particular failure. After years of security remediation, it knows which fixes create regressions.

A stronger Claude or Codex then becomes an input to the product rather than a competitor that destroys it.

This is probably where we would look first if we wanted to start an AI coding company today: find a painful software job with a clear finish line, then use coding agents to own the whole job.

Will human attention become the scarce resource in AI coding?

Human attention already looks like the scarce resource, and this may create a completely new class of developer tools.

Stack Overflow found that most technologists using agents still keep them under supervision. DORA found that some of the time saved during code generation moves into auditing and verification. Meanwhile, coding platforms are making it easier to launch agents asynchronously and let several pieces of work run at once.

That changes the developer's job. A senior engineer may spend less time typing and more time deciding what work should happen, checking whether agents understood the request and approving changes that could affect production.

Current review interfaces were designed for a world where another human wrote the pull request. They assume that reading the diff is a reasonable way to understand what happened.

That becomes painful when ten agents each produce large changes before lunch.

There is room for a different review interface: show the original request, explain which parts of the system changed, run the important tests, surface unresolved uncertainty and give the reviewer enough evidence to make a decision without reading every generated line.

The company that reduces the amount of senior-engineer attention required per safe software change could create more value than another product that generates code 20% faster.

So what’s still open in AI coding?

AI coding is still wide open downstream of generation. If we were starting a company today, we would spend much more time on making autonomous software work trustworthy than on making another assistant write code faster.

The market has already given us several clues. Claude Code, Copilot, Cursor and Cognition have reached enormous scale, making horizontal generation much harder to attack. At the same time, companies such as CodeRabbit and Blacksmith are growing quickly around validation, while Cognition's enterprise usage suggests that migrations and maintenance can support very large agent businesses. Private deployment and AI cost management have also become concrete enterprise problems rather than theoretical ones.

Our strongest category is end-to-end software verification: systems that understand what an agent was supposed to accomplish, test whether it actually accomplished it and produce enough evidence for a company to trust the change.

Legacy modernization comes very close behind. It has huge existing budgets, years of accumulated backlog and much clearer success criteria than generic developer productivity.

Secure agent execution, vertical maintenance agents, private AI engineering and agent-cost control also look genuinely open. Production infrastructure for vibe-coded applications could become another large category as software creation spreads beyond engineering teams.

We would be cautious around another horizontal coding agent, another generic AI IDE, another basic pull-request reviewer or a standalone multi-agent dashboard. Those markets are either already dominated or moving rapidly into platform features.

The opportunity in AI coding these days comes from a simple change in scarcity. Generating code used to consume most of the effort. Now that code is much easier to produce, trust, context, testing, security and human attention become more valuable.

Opportunity	How open is it now?	Our view
End-to-end software verification	Very open	Possibly the clearest new infrastructure layer created by coding agents
Legacy modernization and migrations	Very open	Huge backlog, measurable outcomes and strong enterprise budgets
AI testing and autonomous validation	Very open	More generated code directly increases testing demand
Secure agent execution and permissions	Very open	Agent autonomy creates a new security surface
Vertical maintenance agents	Very open	Easier to measure and defend than a general assistant
Private enterprise AI coding	Open	Strong fit for sensitive and regulated software
AI coding FinOps and model routing	Open	Multi-model usage is creating a new cost-management problem
Production infrastructure for vibe-coded apps	Open	Creating apps is getting easy; operating them still is not
Organization-specific code governance	Open	Internal standards and historical context remain hard to automate
Generic AI code review	Partly open	Demand is strong, but competition and bundling are increasing fast
Generic multi-agent orchestration	Weak	GitHub and developer platforms are already absorbing it
Another horizontal vibe-coding app	Weak	Lovable, Replit and Base44 already have major scale
Another AI-native IDE	Very weak	Coding is moving across IDEs, terminals, repositories and background agents
Another general coding agent	Very weak	Claude Code, Codex, Copilot and Cognition set an extremely high bar

If you want more recent data on this point, please see our latest AI code assistant 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.