Are AI coding agents just a temporary trend?

Can AI coding agents die?

No, AI coding agents will not die. They have already crossed the adoption threshold where they become infrastructure, not a gadget.

The clearest signal is the massive distribution. GitHub Copilot has passed 20 million users across 77,000 enterprises, and GitHub says more than 90% of Fortune 100 companies now use its platform. That is enterprise software penetration at the level where tools get embedded into everything.

The second signal is behavioral. Stack Overflow’s 2025 survey says 80% of developers now use AI tools in their workflows. GitHub’s Octoverse says 80% of new developers on GitHub use Copilot in their first week. So, AI assistance is also becoming part of the default learning and working environment for new software builders.

The third signal is supply-side competition. Microsoft/GitHub, Anthropic, OpenAI, Google, Cursor, Devin/Cognition, JetBrains, GitLab, Amazon, and others are all building into the same wedge: agents that can read a repo, plan changes, edit code, open PRs, run tests, and receive review. Categories can die when only startups believe in them. This one is being institutionalized by the companies that control developer workflows.

The better question is not whether AI coding agents die. It is which form dies. This is the tension we will resolve on this page.

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

Are AI coding agents the future of coding?

Yes. AI coding agents are clearly where coding is going.

Not because developers stop coding, but because the job changes. You spend less time writing every line yourself and more time breaking work into tasks, checking the output, writing tests, and making sure the whole thing fits the system.

You can already see it in the products. GitHub’s coding agent can take a task, work in its own environment, and open a draft pull request. Copilot CLI is turning into a full agentic coding environment that can plan, build, review, and remember across sessions. Google’s Jules is built to read a codebase and work through tasks on its own, not just suggest the next line.

That is the big shift. AI is moving from autocomplete into the actual delivery flow.

Is this just used by amateurs, or do companies trust AI coding agents too?

No, this is not mainly an amateur phenomenon. Enterprises are already one of the strongest adoption signals.

GitHub Copilot’s 77,000-enterprise footprint is the cleanest proof. Accenture’s enterprise study found that over 80% of participants successfully adopted Copilot, 67% used it at least five days per week, and 81.4% installed the IDE extension the same day they received a license. That is unusually fast enterprise uptake for a developer tool, especially one touching source code.

The enterprise trust signal is also visible in scale deployments. Microsoft announced partnerships with Cognizant, Infosys, TCS, and Wipro that collectively exceed 200,000 Copilot licenses.

What about senior engineers? Do they trust AI agents?

Senior engineers do trust AI coding agents, but they trust them differently from juniors.

Juniors often use them to unblock and learn. Seniors use them to compress routine implementation, explore unfamiliar APIs, generate tests, refactor boring code, and draft PRs they still own.

The best evidence is the 2025 RCT on experienced open-source developers: these were not beginners, they averaged five years on mature projects, and they still expected AI to save 24% of time. The surprise was that, in that specific setting, AI slowed them by 19%.

That tells us seniors are using the tools, but their bar is higher: unfamiliar generated code can become review debt.

Are developers becoming dependent on AI coding agents, or are they still optional?

Formally optional, but practically they are becoming hard to avoid in many developer workflows.

The strongest dependency signal is frequency. As seen before, Stack Overflow says 80% of developers use AI tools in their workflows. In Accenture’s Copilot enterprise study, 67% of respondents used Copilot at least five days per week. GitHub says 80% of new developers use Copilot in their first week. Once a tool is used weekly or daily, it stops being a side experiment and starts shaping habits.

The second signal is learning behavior. Stack Overflow reported that 44% of developers learned with AI-enabled tools in 2025, up from 37% in 2024.

Do coding agents save real time, or just create more code to review?

They save real time on the right tasks, but they can absolutely create review debt when used on the wrong tasks.

The positive evidence is strong in enterprise telemetry. In Accenture’s Copilot study, developers saw an 8.69% increase in pull requests, a 15% increase in pull request merge rate, and an 84% increase in successful builds.

Developers also retained 88% of Copilot-generated characters in their editor. That suggests Copilot was not only producing disposable suggestions; a large share of generated code survived into the workstream.

But the negative evidence is equally important. The 2025 randomized trial on experienced open-source developers found that allowing AI tools increased task completion time by 19%, even though developers predicted a 24% speedup and later felt they had saved 20%.

That is the trap: AI can feel faster while making the total cycle slower, especially when the developer must understand a mature codebase, maintain strict standards, and debug subtle generated mistakes.

So, clearly, “lines of code produced” is now a bad productivity metric. The better metric is accepted, tested, merged change per unit of reviewer attention.

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

Is AI-generated code correct, secure, maintainable, and mergeable today?

Today, AI-generated code is often mergeable, sometimes maintainable, usually useful, but not reliably secure or correct without human and automated review.

Mergeability is already proven in real workflows. Accenture reported a 15% increase in pull request merge rate and an 84% increase in successful builds with Copilot. A mobile open-source study found AI-authored PR acceptance rates of 71% for Android and 63% for iOS across verified repositories. AIDev collected 932,791 agentic pull requests across 116,211 repositories, which shows agent-written PRs are no longer theoretical.

Correctness is improving quickly, but benchmark numbers need careful interpretation. SWE-bench Verified is useful because it uses real GitHub issues and test execution, but it is still a 500-task benchmark, not a full production guarantee. Newer harder benchmarks show the gap: Web-Bench reports Claude 3.7 Sonnet at only 25.1% Pass@1 on sequential web-development projects designed by experienced engineers, and SecureAgentBench found the best tested agent setup reached only 15.2% correct-and-secure solutions.

Security is the weakest point. A 2025 large-scale study of 7,703 AI-attributed files found 4,241 CWE instances across 77 vulnerability types, although 87.9% of AI-generated code had no identifiable CWE-mapped issue. Another security benchmark found agents can produce functionally correct code while still introducing vulnerabilities. GitHub Copilot code review also failed to detect critical vulnerability classes like SQL injection, XSS, and insecure deserialization in a 2025 evaluation.

Maintainability is mixed. AI is good at generating readable local code, tests, and documentation, but weak at long-term architectural taste unless the repo has strong conventions and the task is narrow. The under-discussed risk is not that AI writes ugly code once. It is that AI can multiply mediocre patterns across a codebase faster than reviewers can detect architectural drift.

AI code can be merged today, but it should not get a trust shortcut.

Do AI coding agents cost a lot?

Yes, AI coding agents get expensive when they work a lot. It can easily go to $100 per session.

It can be cheap if it’s a short coding exchange: ask for a function, edit one file, generate a small test. At current frontier API prices, that can be under $1 if the session uses roughly 100k input tokens and 20k output tokens. On Claude Sonnet 4.6, that is about $0.60. On GPT-5.3-Codex, about $0.46. On GPT-5.5, about $1.10.

The expensive version is a real agentic session: read the repo, search files, run tests, fail, inspect logs, patch again, summarize, open a PR. At 3M input tokens and 500k output tokens, the same session becomes roughly $12 on GPT-5.3-Codex, $16.50 on Claude Sonnet 4.6, $27.50 on Claude Opus 4.8, and $30 on GPT-5.5. Push that to 10M input tokens and 2M output tokens, and we are around $45, $60, $100, and $110 respectively.

A recent academic study on SWE-bench Verified found agentic coding tasks can consume 1,000x more tokens than code chat or code reasoning, and runs on the same task can differ by up to 30x in token usage.

That means two developers can ask for what feels like the same thing and create wildly different bills.

Have recent events shown coding agent costs can explode?

Yes, 2026 already gave us several clean examples of coding agent cost blow-ups.

The pattern is clear: cost explodes when three things happen together. First, the tool becomes useful enough that developers use it daily. Second, the workflow shifts from autocomplete to autonomous multi-step work. Third, pricing moves from “seat” to “meter.” That is exactly what is happening now.

The AI coding market is entering its cloud-computing moment. Early cloud felt like cheap infinite compute until teams woke up to runaway AWS bills.

Coding agents are following the same curve, but faster, because one employee can now spin up hundreds of dollars of model work without opening a cloud dashboard.

Recent signal	What happened	What it tells us
GitHub Copilot billing change	GitHub moved Copilot to token-based GitHub AI Credits on June 1, 2026, because a quick chat and a multi-hour autonomous coding session could previously cost users the same while GitHub absorbed the inference cost.	Flat-rate pricing broke once agentic usage became normal.
GitHub user backlash	Users reported burning through monthly credit allowances in days, with some cases described as several-hundred-dollar projected bills or up to 100x increases.	The same subscription can feel cheap or expensive depending on task length and model choice.
Anthropic Claude Code estimate reset	Anthropic reportedly doubled Claude Code cost estimates: average enterprise developer cost moved from about $6 to $13 per active day, and the 90th percentile from under $12 to under $30 per active day.	Better models do not always make usage cheaper, because developers run more ambitious tasks.
Uber	Uber reportedly exhausted its 2026 AI budget by April and introduced a $1,500 monthly cap per employee per AI coding tool.	Enterprise adoption can outrun finance planning faster than procurement cycles can react.
Microsoft internal shift	Microsoft reportedly moved engineers away from most Claude Code licenses toward GitHub Copilot CLI by June 30, 2026.	Large firms are starting to consolidate usage where they control distribution and cost.
OpenClaw	Around 100 autonomous Codex agents reportedly consumed $1.3M of OpenAI API tokens in 30 days, across 603B tokens and 7.6M requests.	If agents are allowed to run like cloud workers, the bill looks more like infrastructure spend than software subscription spend.
Linux Foundation Tokenomics Foundation	The Linux Foundation announced a Tokenomics Foundation to create AI cost-management standards.	The industry now sees token economics as a FinOps category, not a niche developer concern.

What costs can we expect tomorrow?

Costs should fall per unit of intelligence, but not enough to make heavy agent usage feel free.

The important distinction is price per token versus tokens per task. Token prices can fall while total bills rise if agents consume far more tokens.

That is already the lesson from 2026: the bottleneck is not only the sticker price of 1M tokens; it is the number of tokens an agent burns to finish one useful change.

The most likely future is not “everything becomes free” but "cheap agents for routine work, expensive agents for serious autonomous work."

Timeline	What likely gets cheaper	Cost impact vs today	Why
2026	Better routing to mini, flash, and specialized coding models.	20–60% lower for routine tasks.	Many agent steps do not need the strongest model; planning, grep, summarization, and test-writing can be routed down.
2026–2027	Prompt caching and repo memory.	30–90% lower on repeated repo context.	Cached input can cost roughly 10% of normal input on Claude and Copilot-style pricing tables.
2026–2027	Batch and async work.	About 50% lower where latency does not matter.	OpenAI, Anthropic, and Google all push cheaper batch or flex-style processing.
2027	Smaller coding models trained for tool use.	50–80% lower for common maintenance tasks.	The market will not send every lint fix, doc update, or CRUD task to the frontier model.
2027–2028	Agent orchestration efficiency.	20–50% lower per completed task.	Better planning, fewer blind retries, better test selection, and budget-aware stopping rules.
2028+	Hardware and inference stack gains.	30–70% lower per token, unevenly.	Chips, quantization, speculative decoding, model distillation, and serving optimization keep improving.
2028+	Frontier autonomous work.	Still expensive.	Long-horizon tasks use more context, more tools, more retries, and more verification as models get more capable.

Can Microsoft, Google, or OpenAI give everyone unlimited free coding agents?

No, unlimited coding agents cannot stay unlimited for everyone.

The reason is structural. A coding agent is not a static software feature. It consumes scarce compute every time it thinks, reads context, calls tools, runs a cloud container, generates output, retries, and reviews itself.

OpenAI lists containers at up to $1.92 per 64GB 20-minute session, web search at $10 per 1,000 calls, and frontier model output up to $30 per 1M tokens for GPT-5.5.

Anthropic lists Claude Opus 4.8 at $5 input and $25 output per 1M tokens, with fast mode as high as $10 input and $50 output for Opus 4.8.

GitHub now explicitly converts Copilot token usage into AI Credits, with 1 credit equal to $0.01.

So the answer is no. Microsoft, Google, or OpenAI can commoditize access, but not unlimited intensive usage.

They can give everyone a free tier. They can bundle agent access into paid subscriptions. They can subsidize early adoption. They can make low-end models feel nearly free.

But they cannot let every user run unlimited frontier agents across repositories all day without either losing money, throttling performance, degrading models, or shifting the bill somewhere else.

We can look at what happened with mobile data and cloud compute. Telecoms sold “unlimited” plans, then added deprioritization, fair-use thresholds, and hotspot caps. Cloud providers made compute easy to start, then built budgets, alerts, reserved instances, and FinOps.

AI coding agents are going the same way: access gets cheaper, serious usage gets metered.

That is why commoditization is less simple than people think. The UI can become free. The heavy work cannot.

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

What are companies doing to reduce or absorb these costs?

Companies are moving from “AI adoption” to “AI cost governance.”

The first tactic is usage-based pricing. GitHub’s June 1, 2026 change is the clearest signal: Copilot plans now include AI Credits, and additional usage is charged by token consumption. OpenAI moved Codex to token-aligned pricing in April 2026 across Plus, Pro, Business, Enterprise, Edu, Health, Gov, and teacher plans. Anthropic has also pushed more Claude Code usage toward explicit credits, API rates, or tighter quota management.

The second tactic is budget controls. GitHub explicitly says admins get budget controls, usage visibility, and credit-based governance. Uber’s reported $1,500 monthly cap per employee per AI coding tool is the real-world version of the same move. Enterprises are no longer asking, “Do developers like the tool?” They are asking, “Which teams, tasks, repos, and models create accepted code per dollar?”

The third tactic is model routing. This is where a lot of savings will come from. Use Gemini Flash, GPT mini, Haiku, or internal small models for cheap steps; reserve GPT-5.5, Claude Opus, or top Codex models for hard reasoning. Copilot’s own pricing table makes the economic spread visible: Gemini 3 Flash is listed at $0.50 input and $3 output per 1M tokens, while GPT-5.5 is $5 input and $30 output. That is a 10x difference on input and output.

The fourth tactic is caching and context discipline. Agents are expensive mostly because they keep reading context. Anthropic’s cache hits cost 10% of base input pricing, and GitHub lists cached input at much lower rates across models. That means the best engineering teams will invest in repo maps, stable context packs, embeddings, reusable summaries, and fewer full-repo rereads.

The fifth tactic is replacing “agent freedom” with “agent accounting.” We should expect default spending caps, per-repo budgets, model allowlists, task approval thresholds, automatic downgrade rules, internal chargebacks, and “cost per merged PR” dashboards. This is not anti-AI. It is how AI becomes enterprise-grade.

Do coding agents create more value than they cost?

Yes, coding agents can create more value than they cost, but only when the team measures value per accepted change, not tokens consumed.

A simple ROI example makes the economics clear. Suppose a senior developer costs $150,000 per year fully loaded. At roughly 1,800 working hours, that is about $83 per hour. If an agentic coding session costs $15 and saves 30 minutes of real developer time, the labor value saved is about $41.50. Net gain: about $26.50 before review risk. If the same $15 session saves two hours, the gross labor value is about $166. Net gain: about $151.

But the inverse is also real. If a $30 agent session produces a messy PR that takes a senior engineer 45 minutes to understand and reject, the cost is not $30. It is $30 plus roughly $62 of reviewer time, plus context switching. That is why AI coding ROI is not “cost per token.” It is “cost per accepted, tested, maintainable change.”

The best available productivity signals are mixed, which is exactly why the ROI question matters. As seen before, GitHub’s Accenture study reported an 8.69% increase in pull requests, a 15% increase in merge rate, an 84% increase in successful builds, and 88% retention of Copilot-generated characters. That points to real workflow value.

But, as seen also before, a 2025 randomized controlled trial on experienced open-source developers found AI tools increased task completion time by 19%, despite developers expecting a 24% speedup and later feeling they had saved time. That points to real review and integration debt.

So, coding agents already have positive ROI on narrow, testable, repetitive, and well-scoped work. They have uncertain or negative ROI on ambiguous, architectural, security-sensitive, or legacy-heavy work unless the team has strong tests and review discipline.

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

What is the ROI threshold for AI coding agents?

The ROI threshold is also surprisingly low.

If a $100/month coding tool saves a developer only 1.2 hours per month at an $83/hour loaded cost, it breaks even.

If OpenAI’s own Codex estimate of $100–$200 per developer per month is the right average range, the tool needs to save roughly 1.2–2.4 hours per month for a senior U.S.-style engineer to justify itself on labor time alone.

That is not a high bar. The hard part is proving the saved time is real and not moved into review, debugging, or post-merge defects.

How strong is Microsoft and OpenAI’s AI coding distribution?

Microsoft and OpenAI’s AI coding distribution is not just strong but, actually, structurally terrifying.

Yes, they have more users, but they also sit on the surfaces developers and companies already use before they even decide to “try an AI coding tool.” Microsoft has Microsoft 365, Windows, Azure, GitHub, VS Code, Teams, enterprise identity, procurement, security admin, compliance, and billing.

OpenAI has ChatGPT, Codex, API distribution, consumer mindshare, and enterprise accounts. That is a different kind of power from “we have a better coding interface.”

The numbers make the gap concrete. Microsoft says Microsoft 365 Copilot already has more than 20 million paid seats and is used by 90% of the Fortune 500. Microsoft 365 itself has over 450 million commercial paid seats. Visual Studio and VS Code together have around 50 million monthly active developers. GitHub publicly says more than 150 million people use the platform. OpenAI’s ChatGPT is reported at roughly 900 million weekly active users, and Codex has reportedly reached more than 5 million weekly users after a rapid post-launch ramp.

That means the default path for a junior developer is already biased toward Microsoft or OpenAI.

They will likely touch VS Code, GitHub, ChatGPT, Microsoft 365, Teams, or a corporate identity system before they ever make a deliberate choice to install Cursor, Devin, CodeRabbit, or another specialist tool. In a normal enterprise, the first AI coding agent a developer meets is not necessarily the best one, but the one already approved by IT.

The distribution multiple is brutal. If we compare Microsoft 365 Copilot’s 20 million paid seats to Cursor’s confirmed $1 billion annualized revenue, and translate Cursor into rough seat-equivalents at $20–$40 per month, Microsoft 365 Copilot is already around 5–10x larger on paid-seat distribution.

If we compare Microsoft’s 450 million commercial Microsoft 365 seats to that same Cursor seat-equivalent base, Microsoft has roughly 100–200x more enterprise seat surface. Against an autonomous agent like Devin, where paid usage is more expensive and narrower, Microsoft’s seat surface is likely hundreds to thousands of times larger.

But the scariest part is not even the absolute scale. It is actually the cost of pushing a product. Microsoft can put an agent into the apps, the IDE, the repo, the terminal, the enterprise admin console, the security dashboard, and the procurement renewal.

Cursor or Devin has to earn every install, every security review, every budget line, every champion, every renewal. Microsoft can make AI feel like a feature. Startups have to make it feel like a product worth buying separately.

Why do people pay for Cursor and Devin when Copilot exists?

People pay for Cursor and Devin because Copilot captures default users, while specialist tools capture serious users.

Cursor proves the point. It passed $1 billion in annualized revenue, raised $2.3 billion at a $29.3 billion valuation, and said enterprise revenue grew 100x in 2025 year-to-date.

Devin proves it from the agent side. Cognition, its parent company, recently raised more than $1 billion at a reported $26 billion valuation. Investors and large customers are clearly betting that autonomous software engineering deserves a separate budget from Copilot-style assistance.

Developer tools reward intensity more than bundling. Average employees accept the default. Strong developers switch when a product feels meaningfully faster, understands the repo better, runs tighter agent loops, handles terminal workflows naturally, or keeps them inside a coding-native environment.

This is not new. Microsoft had Teams, but Slack still became a multi-billion-dollar workflow layer. Google had Docs, but Notion built a major business by owning a different knowledge-work primitive. Salesforce dominated CRM, but HubSpot grew past $3 billion in annual revenue by owning a different buyer, workflow, and usability wedge.

Cursor is not “Copilot but prettier.” We should look at it as the AI-native IDE, the developer cockpit. Devin is not better autocomplete; it is more like delegated engineering work: assign a task, let the agent operate, then review the result.

Copilot wins the default. Cursor and Devin win the users who care enough to defect.

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

What becomes free, and what stays monetizable?

The pattern is very clear. In AI coding, anything that looks like a feature becomes free. Anything that looks like responsibility stays paid.

That is the mistake in the “AI coding will be commoditized” argument. It treats all coding assistance as one category. It is not. A suggestion in the editor is not the same product as an agent that opens a PR, fixes a vulnerability, migrates a service, or enforces a company’s architecture rules.

The monetizable layer moves upward. Not “write me code.” More like: “own this task inside our engineering system, follow our standards, respect our security policy, control the bill, and produce something reviewable.”

Part of the AI coding stack	Probability of commoditization	Why
Inline autocomplete	95%	It is already expected inside IDEs, and small models can handle much of it cheaply.
Basic code chat	90%	Chat is now a generic interface across ChatGPT, Copilot, Gemini, Claude, and IDEs.
Simple single-file edits	85%	Model quality is good enough, context needs are small, and bundlers can absorb cost.
Boilerplate generation	90%	Low differentiation, easy to benchmark, easy to route to cheaper models.
Test stub generation	75%	Useful but increasingly standard inside IDEs and CI workflows.
Repo Q&A	65%	More context-intensive than chat, but likely bundled into GitHub, IDEs, and enterprise search.
Multi-file refactors	45%	Still requires repo understanding, test awareness, and rollback discipline.
Pull request review	40%	Code review touches quality, liability, compliance, and team standards, so teams will pay for trust.
Security remediation	25%	Vulnerability fixing has clear ROI, high urgency, and budget from security teams.
Autonomous issue-to-PR agents	25%	Long-running agents consume real compute and need reliability, observability, and governance.
Migration agents	20%	Framework upgrades, cloud migrations, and dependency migrations have large enterprise budgets.
Regulated-industry coding agents	15%	Finance, healthcare, defense, and legal need audit, policy, data controls, and domain-specific workflows.
Agent orchestration and governance	10%	Enterprises will pay to control cost, permissions, model choice, logs, and approval workflows.

Can a new AI coding startup still win tomorrow?

Yes, a new AI coding startup can still win, but not by launching another generic coding assistant.

A new startup has almost 0% chance if it launches “Copilot, but better” as a horizontal autocomplete or chat product. Microsoft, Google, OpenAI, Anthropic, JetBrains, GitLab, Amazon, and Cursor are already too far ahead in distribution, funding, model access, and developer mindshare.

A new startup has maybe 20–30% odds if it attacks a painful workflow with measurable ROI: AI code review, flaky test repair, dependency upgrades, security remediation, data pipeline maintenance, framework migrations, mobile app modernization, CI failure debugging, or legacy code documentation. These are narrower than “coding,” but they have buyers, budgets, and measurable outcomes.

A new startup has more than 50% odds only if the market is not actually “coding assistant” anymore, but a vertical engineering system. For example: an agent for bank compliance code, healthcare integration, SAP customization, insurance claims systems, embedded automotive software, game development pipelines, defense software audits, or large-scale cloud migrations. In those categories, Microsoft’s distribution helps, but it does not automatically understand the workflow.

Is AI coding a winner-take-all market?

No, AI coding is not winner-take-all. It is more like winner-take-default plus specialist-take-margin.

Microsoft, OpenAI, and Google are positioned to own the default layer because they control identity, docs, email, IDEs, repos, cloud, and enterprise procurement. That layer will be massive. It will create enormous usage and put pressure on standalone tools.

But the market is too fragmented to collapse into one winner. A single assistant can be good enough for many tasks, but not trusted for every context.

The strongest comparison is enterprise software. Microsoft Teams did not kill Slack. Google Docs did not kill Notion. Salesforce did not kill HubSpot. The bundled product often wins default distribution, while the specialist survives by owning something else.

Coding agents will follow the same pattern. GitHub Copilot can be the default. Cursor can be the AI-native coding environment. Devin can be the delegated engineering agent. CodeRabbit or Qodo can own AI review. Security-first tools can own remediation. Vertical agents can own regulated or domain-heavy work.

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

Will coding agents become horizontal tools or vertical tools?

Coding agents will start horizontal, but the durable money will probably move vertical.

The first wave had to be horizontal because the basic capability was new: generate code, explain code, edit code, chat with a repo. Horizontal tools were the fastest way to reach developers and prove demand. That is why Copilot, Cursor, ChatGPT, Claude Code, Codex, and Gemini Code Assist matter.

But horizontal tooling gets commoditized once every large platform has similar baseline capability. The next differentiation comes from context: company standards, code ownership, compliance rules, deployment history, service dependencies, security posture, industry regulations, and team conventions. That context is not generic. It is vertical or workflow-specific.

We are already seeing the shape of this. Legal AI company Harvey is embedding domain-specific agents into Microsoft 365 and reportedly reached around $190 million ARR. Code review tools are becoming governance layers, not just comment bots. Security startups are forming specifically because AI-generated code creates new review and vulnerability problems. Meta’s internal research shows AI has increased code output enough that review capacity becomes the bottleneck, which creates a market for specialized review automation.

Our final view: horizontal agents win reach, vertical agents win margin.

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.