Cursor just hit $2 billion ARR in 26 months
Inside Anysphere’s rocket ride from dorm room to $29B valuation, the competitive bloodbath reshaping developer tools, and the unit economics behind it
Cursor’s annualised recurring revenue reached $2 billion in February 2026, according to Bloomberg reporting on 2-3 March. The company crossed $1 billion just three months earlier. Two billion dollars for a desktop application that helps programmers write code.
Stack that against any SaaS company in history. Snowflake took six quarters to travel from $1 billion to $2 billion ARR. Slack took roughly six years to reach $1 billion. Zoom took nine. Shopify, fifteen. Cursor ran from essentially zero (about $1 million ARR in December 2023) to $2 billion in 26 months. Wild.
The valuation tracked accordingly. Anysphere, the parent company, was valued at $29.3 billion in its November 2025 Series D, raising $2.3 billion. That round came five months after a $900 million Series C at $9.9 billion. The company began with a pre-seed of roughly $400,000 in April 2022. From $400,000 to $29.3 billion is a 73,250x increase in 43 months. Even by venture capital standards, absurd.
And the revenue is real. These aren’t bookings or commitments or letters of intent. Cursor is a subscription product. People pay $20, $60, or $200 per month. Teams pay $40 per user. The money clears every billing cycle. By mid-2025, the company already counted the majority of Fortune 500 companies as customers, alongside OpenAI, Stripe, Spotify, Midjourney, and Perplexity.
The developer community is predictably split. One camp sees vindication: they adopted early, and the numbers proved them right. The other asks pointed questions about unit economics, sustainability, and what happens when a company spends every dollar it earns on API calls to someone else’s models. Both sides have a point.
The hockey stick of all hockey sticks
Cursor’s revenue trajectory breaks every recognised SaaS growth pattern. Most B2B companies celebrate each doubling after quarters of grinding. Cursor has been doubling every three months.
The documented progression: $1 million ARR in December 2023. $48 million by October 2024. $100 million by January 2025. $300 million by April 2025, per Michael Truell’s own disclosure to Lenny Rachitsky. $500 million by June 2025. $1 billion by November 2025. And now a staggering $2 billion in February 2026.
A 2,000x increase in 26 months. The doubling time from $1 billion to $2 billion: roughly 90 days.
At $2 billion ARR with around 300 employees, Cursor generates roughly $6.7 million in revenue per employee. Google generates about $1.7 million. Even the most efficient SaaS companies rarely crack $500,000.
The consumer comparison people reach for is ChatGPT, which hit 100 million users in two months. But ChatGPT is a freemium consumer product with far lower average revenue per user. Cursor charges $20 to $200 per month per seat. Growing this fast while charging real money, not just stacking free signups, is what makes the trajectory very unusual.
One metric from early 2025 captures the business dynamics. Cursor reported a 36% freemium-to-paid conversion rate. The industry average for SaaS freemium products sits between 2% and 5%. At 36%, the free tier is less marketing funnel, more brief audition before the cheque gets written.
That conversion rate points to something about the product itself. People try Cursor and, within a short window, decide they can’t go back. The switching cost isn’t contractual. It’s experiential. Once your workflow reshapes itself around an AI-native editor, reverting to plain VS Code feels like giving up autocomplete.
The fundraising escalator
Venture capital noticed early. Anysphere’s funding rounds compressed as aggressively as its revenue milestones.
Total raised: roughly $3.2 to $3.4 billion. The jumps: $400,000 to $400 million to $2.6 billion to $9.9 billion to $29.3 billion. Each round roughly tripled the prior valuation.
When a company raises $2.3 billion five months after raising $900 million, investors are competing fiercely for allocation, revenue is validating every prior bet, and the company sees opportunities to spend capital at a rate that justifies the dilution.
Four kids from MIT
Before Cursor became a financial phenomenon, it was four classmates writing code in an MIT dorm room. Michael Truell, Sualeh Asif, Aman Sanger, and Arvid Lunnemark, all computer science graduates from the class of 2022, founded Anysphere in 2022 with that initial $400,000 (which looks like a rounding error now).
They were ~20 years old. Three and a half years later, all four are billionaires. At the Series D valuation, each co-founder’s estimated stake of around 4.5% was worth approximately $1.3 billion.
The founding story is, in some ways, classically Silicon Valley. Four technical founders, a pre-seed that barely covers rent, an insight the incumbents are too slow to act on. But the insight itself was sharper than the typical startup pitch. Michael and the other three noticed something specific: GitHub Copilot’s plugin architecture was limited in ways that mattered. Copilot operated as a suggestion engine bolted onto an existing editor. It could autocomplete lines. It couldn’t reimagine the editing experience.
The bet: fork VS Code, the most popular code editor in the world, and rebuild it as an AI-native environment. That meant controlling the entire editor experience. The file tree, the terminal, the diff view, the chat, the tab completion. Every surface of the editor became an integration point for language models.
The decision to fork VS Code rather than build a plugin was the most consequential technical choice in the AI coding tools market. It traded compatibility for control.
Frankly, the audacity is easy to underestimate in hindsight. VS Code had (and has) an enormous extension ecosystem. Forking it meant maintaining compatibility with that ecosystem while diverging from it, taking on the maintenance burden of an editor that Microsoft staffs dozens of engineers to develop. Four people in a dorm room decided they could do it better than a trillion-dollar company.
They were right.
The team today
By November 2025, Anysphere had grown to over 300 employees. Still remarkably lean for a $2 billion revenue run rate. Snowflake had over 5,000 employees when it crossed $2 billion ARR. Slack had roughly 2,500 at $1 billion.
That revenue-per-employee figure (roughly $6.7 million at current rates) reflects a company where the product does most of the selling. Cursor doesn’t employ a traditional enterprise sales force. Developers download it, try it, pay for it, and tell their colleagues. The bottoms-up adoption model that Slack pioneered in the 2010s is operating at a pace Slack’s founders probably couldn’t have imagined.
How Cursor actually makes money
The pricing is simple. Four tiers for individuals, two for teams.
Cursor pricing tiers (as of March 2026)
The $20 Pro tier is the volume centre. When Cursor disclosed 360,000 paying customers in early 2025 at $100 million ARR, average revenue per user worked out to roughly $23 per month, tracking closely with the Pro price point. At $2 billion ARR, the paying base has grown substantially, though the company hasn’t disclosed updated numbers.
Look at the jump from Pro to Ultra. A 10x price increase from $20 to $200 suggests heavy usage generates roughly 10x the cost. The tier structure isn’t about feature differentiation. It’s about usage caps: how many completions, how many chat messages, how many agent runs per billing cycle.
The product evolution
Cursor launched as a smarter autocomplete tool. By early 2026, it had evolved into something closer to an autonomous coding agent.
The milestones came fast. In October 2025, Cursor launched a proprietary LLM called Composer, described as 4x faster than previous models and designed to reduce API costs by handling routine coding tasks without calling out to expensive external models.
In February 2026, Cursor shipped what the company called its “most significant update in Cursor’s history”: cloud agents with computer use. Background agents running on virtual machines that can execute multi-step tasks, run tests, browse documentation, and make commits, all without the developer watching. Alongside this came a plugin marketplace with integrations for Figma, Linear, Stripe, and AWS.
The direction is plain. Cursor started by making individual developers faster at typing code. It’s now positioning itself as a platform that handles entire workflows autonomously, from ticket to pull request.
Cloud agents are the logical endpoint of the AI coding tool thesis: the developer becomes a reviewer rather than a writer. Whether that’s a promotion or a demotion depends on who you ask.
The zero-margin paradox
Reports consistently suggest that Cursor spends roughly 100% of its revenue on AI API costs. Every dollar that comes in goes right back out to model providers.
This is the zero-margin paradox at the heart of AI coding tools. The product is demonstrably valuable. Users convert at 36%, an absurd rate. Revenue doubles every quarter. But the core input cost (access to frontier language models) is set by a handful of model providers, and it’s enormous.
The Composer bet
Cursor’s proprietary Composer model, launched in October 2025, is the company’s primary response to this cost problem. By training its own model for routine coding tasks (the bread-and-butter completions and edits that account for most API calls), Cursor can cut per-query costs on common operations while reserving expensive frontier model calls for the hardest problems.
The question is whether a company that has raised $3.4 billion can train models competitive enough to handle the workload. The top AI labs (Anthropic, OpenAI, Google) are spending tens of billions on training. Cursor doesn’t need to match them at the frontier. It needs to match them on the vast middle of coding tasks where a smaller, faster, cheaper model can do the job.
The SemiAnalysis observation that Microsoft is “renting GPUs to the barbarians who will ruin their castle” applies equally to the model providers selling API access to Cursor. Every successful Cursor query is a query that didn’t go through Copilot.
Where the money goes
The roughly $3.4 billion in total funding provides enormous runway, even at zero margins. If Cursor is truly spending all of its $2 billion ARR on costs, the funding covers the gap for years. The bet is that one or more of these things happen before the money runs out:
Composer and successor models get good enough to handle 70-80% of queries at a fraction of the cost
Model API prices keep falling (they have, dramatically, over the past two years)
Enterprise contracts at custom pricing provide higher margins than individual subscriptions
The platform play (marketplace, integrations) creates revenue streams that don’t scale with model costs
Well, the bears point out that every one of those assumptions has a counterargument. Proprietary models require enormous ongoing investment. API prices might fall, but so might subscription prices as competition intensifies. Enterprise contracts take time and sales infrastructure. And the platform play requires an ecosystem that doesn’t yet exist at scale.
The optimistic case is simple: Cursor occupies the position Amazon Web Services held in 2006. Lose money on every transaction, gain market share, build lock-in, then raise prices or cut costs once you own the market. The pessimistic case is also simple: AWS controlled the infrastructure. Cursor is renting it.
The great IDE war of 2026
The AI coding tools market has consolidated into something approaching three-way parity, and Cursor is in the mix against competitors with vastly deeper pockets. According to Recon Analytics data from January 2026, preference share among paid AI coding subscribers in the US breaks down to roughly GitHub Copilot at 24.9%, Cursor at 24%, and Claude Code at 24%. A year earlier, Copilot’s dominance was unquestioned.
GitHub Copilot: the incumbent under pressure
Microsoft’s Copilot disclosed 4.7 million paid subscribers in January 2026 earnings, putting it north of $1 billion in ARR with over 20 million cumulative users. By any normal standard, enormous numbers.
But the trend tells a different story. Copilot’s preference share among US paid AI subscribers dropped from 18.8% to 11.5% between mid-2025 and January 2026, according to Recon Analytics. That’s meaningful erosion, even if total subscribers still grow.
Honestly, there’s a telling corporate signal buried in Microsoft’s earnings. The company stopped reporting AI revenue as a separate line item after disclosing $13 billion in AI run-rate in January 2025. Companies stop breaking out metrics that stop improving. When a number accelerates, you shout about it. When it decelerates, you fold it into a larger segment where it’s harder to isolate.
Claude Code: the model provider as competitor
Anthropic’s Claude Code reached general availability in May 2025 and has grown aggressively since. Sacra estimates the Claude Code division generates approximately $2.5 billion in ARR, though the boundary between Claude Code revenue and broader Anthropic API revenue is blurry.
Claude Code takes a different approach. It’s a command-line tool, not an IDE. It doesn’t try to replace your editor. It operates alongside whatever environment you already use, including Cursor itself (many developers run both). The competitive dynamic is complicated by the fact that Cursor has historically relied on Claude as one of its primary underlying models.
The Windsurf saga
The most instructive competitive episode of 2025 involved Windsurf, the AI coding editor that briefly looked like a serious challenger.
Windsurf was doing roughly $82 million in ARR when OpenAI attempted to acquire the company for approximately $3 billion in May 2025. That deal collapsed. Google licensed Windsurf’s technology in July 2025. Cognition acquired its intellectual property and team later that same month.
Three months. From $3 billion acquisition target to absorbed for parts. This market doesn’t have room for fourth place.
The broader market
The total prize is considerable. The AI coding tools market was valued at roughly $7.37 billion in 2025, with projections ranging from $45 billion to $127 billion by 2032 depending on methodology. That wide range reflects genuine uncertainty about adoption curves, pricing evolution, and whether these tools expand the market for software development or just redirect existing spending.
Worldwide AI spending is forecast to reach $2.52 trillion in 2026, a 44% year-over-year increase according to Gartner. Coding tools are a thin slice of that total, but they’re the slice that touches the people building everything else.
Why Cursor won (so far)
Three factors explain Cursor’s advantage, and none of them are the ones that surface-level takes focus on.
The fork advantage
Forking VS Code rather than building a plugin gave Cursor something no plugin-based competitor can replicate: control over the entire editing experience. When Cursor wants to add inline diffs, a composer panel for natural language instructions, or background agents in cloud VMs, it doesn’t work within the constraints of someone else’s extension API. It modifies the editor itself.
This sounds like a minor technical distinction. In practice, it’s an enormous competitive moat. Every plugin-based tool (including Copilot, unless Microsoft decides to fork its own editor, which would be an extraordinary admission of failure) is limited by what the host API allows. Cursor’s features are limited only by its own engineering capacity.
The model-agnostic bet
Cursor doesn’t build frontier models. Composer handles routine tasks, but the heavy lifting still goes to Claude, GPT-5, and others. This means Cursor benefits from improvements in any frontier model, from any provider. Anthropic ships a better Claude, Cursor gets better. OpenAI ships a better GPT, same thing.
Copilot, by contrast, is tightly coupled to OpenAI. It benefits from OpenAI improvements but can’t easily switch if Claude or Gemini pull ahead on coding benchmarks. Cursor’s model-agnostic architecture means it’s always running the best available model for each task, creating a subtle but persistent quality edge.
The developer-first distribution
Cursor spread through teams the way Slack spread through companies a decade ago: one person tries it, becomes visibly more productive, and the rest start asking what they’re using. The 36% freemium-to-paid conversion rate reflects a product that sells itself through direct experience.
Cursor’s distribution inverts the traditional enterprise sales playbook. Instead of selling to CTOs who push tools down to developers, Cursor sells to individual developers and lets adoption pressure flow upward.
By mid-2025, Cursor reported over 100,000 new projects being built daily on the platform. That organic activity creates a flywheel: more users, more feedback, better product, more users. The customer list (OpenAI, Stripe, Spotify, Midjourney, Perplexity) serves as social proof that accelerates adoption among the next tier of companies.
There’s also, frankly, a demographic factor. Cursor was built by people who were in their early 20s when they started. The product reflects the sensibility of developers who grew up with language models as a baseline assumption, not a novelty. The design choices, the interaction patterns, the defaults: they feel native to a generation that expects AI assistance as a given. Older tools, including Copilot, feel retrofitted.
The productivity question beneath the hype
About 46% of code written by active developers now comes from AI, according to 2026 estimates. Nearly half of all production code, at least in first draft, generated by a machine. The number keeps climbing.
The assumption baked into Cursor’s valuation (and the broader AI coding market) is that AI tools make developers substantially more productive. More productive developers mean companies need fewer of them, or can build more with the same headcount, or can ship faster. Any of those outcomes justifies $20-200 per developer per month.
But the evidence is more complicated than the narrative.
The MERL Tech study
A randomised controlled trial by MERL Tech found that experienced open-source developers were actually 19% slower when using AI coding tools compared to working without them. The study has been disputed on methodological grounds (the tasks may not have been representative of typical professional work, and the sample size was modest), but it’s the most rigorous experimental evidence available, and it points in an uncomfortable direction.
The result deserves consideration, not dismissal. One plausible explanation: experienced developers have deeply optimised workflows, and inserting an AI assistant disrupts those workflows in ways that cost more time than the AI saves. The developer has to review generated code, decide whether to accept or reject it, mentally model what the AI is doing, and handle the cases where it’s wrong. For someone who already knows the codebase and types fast, that overhead may exceed the benefit.
The MERL Tech finding doesn’t mean AI coding tools are useless. It means productivity gains may be concentrated among less experienced developers and on certain kinds of tasks, while experienced developers pay a “supervision tax” that eats into the time savings.
The Karpathy paradox
Andrej Karpathy, the former Tesla AI director who coined the term “vibe coding,” has himself stopped vibe coding for serious projects. When the person who named the practice stops doing the practice, that tells you something.
The likely explanation is familiar to anyone who’s used these tools extensively. AI-generated code works brilliantly for greenfield projects, prototypes, scripts, and throwaway work. It works less well for production systems with complex state management, subtle performance requirements, and codebases carrying years of institutional knowledge. The gap between “write me a function that does X” and “modify this function in a way that respects the 47 implicit constraints in this codebase” is precisely where AI coding tools struggle most.
The dependency question
Critics raise a deeper concern: AI coding tools create a dependency that gradually erodes fundamental programming skills. If 46% of your code is AI-generated and you’re spending your time reviewing rather than writing, you’re exercising a different cognitive muscle. Over time, the review muscle strengthens and the creation muscle atrophies.
Sure, this argument has been made about every productivity tool since pocket calculators. And historically, the alarmists have mostly been wrong. Calculators didn’t make mathematicians worse at maths. Spell-checkers didn’t make writers worse at spelling. (Well, maybe a little.)
But code is different from arithmetic or spelling in one respect. Code involves holding complex systems in your head, and the ability to do that comes from practice. If AI handles the routine work that builds and reinforces that mental model, the question is whether the model degrades.
Nobody has a definitive answer yet. The question is worth more than the dismissive hand-waving it usually gets.
What $2B means for the profession
There are roughly 50 million software developers worldwide, according to SlashData’s 2025 estimates. Cursor alone generates $2 billion in annual revenue from a fraction of them. Add Copilot’s billion-plus, Claude Code’s estimated $2.5 billion, and the long tail of smaller competitors, and the AI coding tools market is already extracting serious money from the global developer population.
Where does it come from? If companies pay $20-40 per developer per month and those tools genuinely boost productivity, the cost is easily justified. A developer earning $150,000 per year costs their employer about $200,000 fully loaded. A 10% productivity improvement from a $480/year tool is an extraordinary return.
But if the productivity improvement is real and large, the second-order effects matter more.
The junior developer squeeze
The tasks AI coding tools handle best (boilerplate, simple functions, CRUD operations, test generation, documentation) are exactly the tasks junior developers have historically cut their teeth on. If AI increasingly handles those, the entry-level pipeline narrows.
This is already showing up anecdotally. Hiring managers report that junior candidates who leaned heavily on AI tools during education struggle with basic debugging and problem decomposition when the tools aren’t available. The tools that make senior developers more productive may simultaneously be making the path from junior to senior harder to walk.
The IDE as platform
Cursor’s February 2026 plugin marketplace signals ambition beyond code completion. If Cursor becomes the environment where developers spend their entire working day, it becomes a platform in the same sense as iOS or Android. Third-party developers build on it. Companies integrate with it. The editor becomes the operating system for software development.
The integrations tell the story. Figma for design-to-code. Linear for project management. Stripe for payments. AWS for deployment. If all of these are accessible within the Cursor interface, a developer might never leave the editor. That’s lock-in of the most powerful kind: not contractual, but habitual.
The 85% question
One statistic hangs over the conversation: 85% of developers use the JetBrains IDE ecosystem (IntelliJ, PyCharm, WebStorm, and others). That’s a massive installed base that Cursor, a VS Code fork, doesn’t natively serve. JetBrains has its own AI assistant, and many of these developers have years of muscle memory and configuration invested in their setup.
Right, Cursor’s growth so far has come predominantly from the VS Code universe. There’s an enormous population of developers who haven’t considered switching because the cost, measured in workflow disruption, isn’t worth it. If Cursor can crack the JetBrains population, the growth story has another gear. If it can’t, there’s a ceiling.
The bull case and the bear case
Any honest analysis has to hold two contradictory ideas at once: this is the fastest-growing B2B software company in history, and it might be spending every dollar it earns.
The bull case
The bull case rests on historical parallels and trend lines.
Model costs are falling. Cost per token for frontier language models has declined by roughly an order of magnitude per year since GPT-3. If that continues (and there’s structural reason to think it will, as hardware improves and architectures get more efficient), Cursor’s margin problem resolves itself over time. The company that can survive at zero margins while costs fall becomes enormously profitable when costs fall far enough.
Proprietary models will absorb volume. Composer and its successors don’t need to match Claude or GPT-4 on the hardest tasks. They need to handle the 70-80% of queries that are routine enough for a smaller model. If Cursor shifts that traffic onto its own infrastructure, the unit economics improve dramatically.
The platform play creates new revenue. A marketplace with Figma, Linear, Stripe, and AWS integrations can charge commissions, listing fees, or premium placement. These revenue streams don’t scale with model costs.
The market is still early. With ~50 million developers worldwide and AI coding tool penetration still in the single-digit millions of paid seats, the addressable market is enormous. Projected growth to $45-127 billion by 2032 would make Cursor’s current $2 billion look like an early milestone.
Network effects are compounding. More users generate more data about coding patterns, which improves autocomplete and agent quality, which attracts more users. The company with the most users builds the best product. This flywheel favours the leader.
The bear case
The bear case is serious too, and it doesn’t require any bull case assumptions to be wrong. It just requires them to be slow.
The margin problem may not self-correct. If Cursor grows usage faster than model costs decline, margins stay at zero or go negative. A company generating $2 billion in revenue with zero margin is functionally a non-profit that has raised $3.4 billion in donations from venture capitalists.
Model providers are competitors. Anthropic (Claude Code) and potentially OpenAI both sell AI coding tools directly. They have a structural cost advantage: they don’t pay API costs to themselves. Cursor is funding the R&D of one of its principal competitors every time it calls Anthropic’s API.
The fork is a liability, not just an asset. VS Code is actively developed by Microsoft, which ships updates constantly. Cursor has to merge upstream changes while maintaining its divergent features. As VS Code evolves (and as Microsoft potentially adds AI-native features directly), the maintenance burden grows. If Microsoft breaks compatibility in ways that hurt forks, Cursor’s entire foundation is at risk.
Enterprise stickiness isn’t proven. Cursor’s bottoms-up adoption is powerful, but enterprise procurement departments evaluate vendors on SOC 2 compliance, data residency, long-term viability, and vendor lock-in. A company with zero profits and a dependency on third-party model providers may struggle to pass enterprise screens at the largest organisations.
Copilot has distribution money can’t buy. Every GitHub repository, every Azure subscription, every Visual Studio installation is a potential Copilot distribution point. Microsoft has a sales force covering every major enterprise on Earth. Cursor’s bottoms-up model is fast, but Microsoft’s top-down model is relentless.
The honest assessment
The best comparison isn’t other SaaS companies. It’s the early days of cloud computing.
In 2006, Amazon Web Services was a money-losing experiment most analysts dismissed. Infrastructure costs were enormous, margins thin, and the incumbents (IBM, Oracle, HP) had distribution advantages that seemed insurmountable. AWS succeeded because it was right about the direction of the market and willing to sustain losses long enough for the economics to tip.
Cursor’s bet is structurally similar. The market direction (developers writing code with AI assistance) seems irreversible. The question is whether Cursor can sustain the economics long enough, and whether the competitive dynamics are forgiving enough, for the bet to pay off.
The $3.4 billion in funding provides runway measured in years, not months. The $2 billion in revenue demonstrates product-market fit at scale. Near-parity with Copilot and Claude Code demonstrates competitive viability.
But the margins. The margins remain the thing nobody on the bull side wants to discuss in concrete terms, and the thing the bear side keeps circling back to.
Cursor may be the fastest-growing B2B software company in history. It may also be the largest company ever to generate $2 billion in revenue without demonstrating a path to profitability. Both things can be true at once.
The four founders who started in an MIT dorm room in 2022 are now billionaires running a company that has raised more capital than most public companies ever will. Their product has reshaped how millions of developers write software. Their revenue growth has shattered records that stood for decades.
And every morning, they face the same question every AI application company faces in 2026: can you build a durable business when your most important input (access to intelligence) is a commodity controlled by someone else?
The next twelve months will provide an answer. The proprietary models will either work well enough to shift economics, or they won’t. The marketplace will either generate independent revenue, or it won’t. The enterprise contracts will either provide premium margins, or they won’t.
Anthropic’s overall revenue reportedly sits at $14 billion ARR as of February 2026. OpenAI closed 2025 at over $20 billion ARR. The model providers are getting richer at least as fast as the application layer. Whether the application layer captures durable value, or merely intermediates it before the model providers absorb the market, is the defining question of this moment in AI.
Cursor’s $2 billion is either the beginning of that answer or a very expensive way of asking the question.
[1] The $6.7 million revenue-per-employee calculation uses the $2B ARR figure and the most recent headcount disclosure of 300+ employees from November 2025. Actual current headcount may be higher, which would reduce this figure.
[2] Market share figures from Recon Analytics measure “preference share among US paid AI coding subscribers,” which is a narrower metric than global market share by revenue or total users. Different methodologies produce different rankings.
[3] The MERL Tech RCT measured task completion time on specific open-source contribution tasks. Critics have argued that these tasks are not representative of typical professional development work, which involves more context-switching, codebase navigation, and collaborative workflows where AI tools may provide different benefits.