Between them, the two companies that dominate AI-assisted software development are worth close to 90billion,morethanseveralenterprise−softwaregiantstheyarequietlydisplacing.Cursor,builtbyAnysphere,doubleditsannualizedrevenuetoabout2 billion in three months and is negotiating a round of roughly 2billionatavaluationabove50 billion, co-led by Andreessen Horowitz and Thrive Capital with NVIDIA joining, Bloomberg, TechCrunch, and CNBC reported in April 2026. Cognition, the startup behind the autonomous coding agent Devin, closed 1billionata26 billion valuation in May and is now in talks that would value it near $40 billion.
The growth numbers have no precedent in business software. By most tallies, Cursor became the fastest-scaling B2B company on record, ahead of Slack and Zoom at the same stage. Devin, for its part, writes roughly 90 percent of its own code, a fact that sounds like a marketing line but turns out to be load-bearing, because it is precisely this headcount-free scaling that makes the companies so valuable on paper, and so strange to staff in practice. They reached tens of billions in value with teams small enough to fit in a single floor, which leaves no internal bench to promote when revenue suddenly demands a full organization.
A Duopoly No One Planned For
None of this was the plan eighteen months ago. The category was an afterthought, a code-completion plugin bolted onto an editor. Gartner’s August 2025 forecast, that 40 percent of enterprise applications would embed task-specific AI agents by the end of 2026, up from under 5 percent in 2025, read at the time like a stretch target. It now looks conservative.
The money followed the usage, and the usage followed the talent. In July 2025, Google paid $2.4 billion to hire Windsurf’s chief executive Varun Mohan, co-founder Douglas Chen, and dozens of engineers outright, a deal that ended OpenAI’s own months-long attempt to buy the company. The object of that transaction was not a product. It was the people.
A year later, the same logic governs the entire sector. When a model company needs capability, it can buy compute. When it needs distribution, it can buy marketing. When it needs the handful of people who actually know how to make an autonomous system behave in production, it has to buy them one at a time, usually from a rival. The scarcest input is no longer the model, and it is no longer the capital. It is the humans who can turn a capable model into a product a business will pay for. SunTzu Recruitment’s industry advisor, who counsels founders through their first engineering hires, put the shift in starker terms: two years ago the questions were all about model benchmarks; today the first question is whether a team can be assembled at all.
Demo Is Easy. Production Is the Wall.
Here is the uncomfortable number sitting beneath all the euphoria. Gartner and IDC data published in August 2026 found that 89 percent of enterprise AI agent pilots stall before reaching production. The demo works. The deployment does not.
A recruitment specialist at SunTzu Recruit put it plainly: building an agent that does a task correctly once is a weekend project. Building an agent that does the same task correctly ten thousand times a day, against a customer’s real data, without burning the customer’s trust or budget, is a full engineering organization. Most companies have the first. They do not have the second.
What “production” actually means is where the résumé breaks down. It is evaluation harnesses that catch when the agent quietly changes its own behavior. It is guardrails that stop a coding agent from committing a mistake into a customer’s repository at two in the morning. It is the unglamorous work of latency budgets, fallback paths, and permission boundaries that a demo never exposes, and most of it never shows up in a job posting, which is part of why the searches drag on. A SunTzu Recruit consultant who covers the sector observed that candidates who can talk fluently about model architecture outnumber, by a wide margin, the ones who can describe how they kept a system alive at scale.
That gap is where the hiring crisis lives. ManpowerGroup’s 2026 talent survey ranked AI-related skills as the world’s most severe shortage, above every other discipline it tracks. The average base salary for an AI engineer in the United States reached 206,000in2025,uproughly50,000 in a single year. In San Francisco, an AI agent engineer’s total compensation, base plus equity plus bonus, runs around $252,000 at the median, according to 2026 compensation data. Salaries are climbing, and offers are still not closing.
The Roles Nobody Trained For
The specific people in demand did not exist in this form three years ago, and almost no one trained for them. A senior consultant at SunTzu Recruit named the three hardest searches. First, agent engineers who can design reliable multi-step systems rather than single-shot prompts. Second, deployment specialists who can carry a working prototype through an enterprise’s security review, data-governance checks, and production rollout. Third, the commercial leaders who can sell a product whose value a buyer has to be walked through twice.
A partner at SunTzu Recruitment added a detail that surprises outsiders: the commercial hires are often harder than the technical ones. The companies that reached revenue fastest did so with founders who could demo their own product. Replacing that with a go-to-market team that can sell to a chief information officer, not a developer, is a completely different skill, and the international market has almost no bench for it.
That is why AI agent platforms and the AI agent infrastructure companies beneath them increasingly turn to specialized search rather than job boards. SunTzu Recruitment has begun structuring these searches around a single question, whether the candidate has ever shipped an agent past a pilot, because the résumé keyword almost never predicts the answer. Searches the firm runs for this cross-border talent pool now routinely stretch six months or more, and overseas clients report the same pattern from Singapore to London. The true unit of artificial intelligence readiness, it turns out, is people, not models.
The Bottleneck Moves Upstream
To be fair, not every company is chasing the same handful of engineers. A different picture emerges in the thousands of enterprises that are not building agents but buying them. There, the shortage is less about engineers and more about the internal operators who can own an agent after the vendor walks away. A Shenzhen-based firm in the cross-border e-commerce space, for instance, spent two quarters looking for someone who could run its newly purchased coding tool across a team of forty, and settled for a compromise hire. The pattern repeats: the technology arrives faster than the operating layer that surrounds it.
That said, the direction of travel is identical everywhere. The binding constraint on the AI coding boom has stopped being the code, and become the people who ship it. A practice lead at SunTzu Recruitment expects the shortage to widen before it narrows, because universities cannot mint agent engineers as fast as the funding arrives. Long story short, whoever can close that gap first wins the next decade.
Sources: Bloomberg / TechCrunch / CNBC via Idlen and LetsDataScience (April 2026 — Cursor/Anysphere in talks to raise ~2billionat>50 billion valuation, co-led by a16z and Thrive Capital with NVIDIA; ~2billionARR);NoqtaandIdlen(May2026—Cognition/Devinraises1 billion at 26billion);VibecodedThis(August2026—Cognitionintalksat 40 billion); Computerworld, The Information, and StartupMap (July 2025 — Google’s 2.4billionhireofWindsurfCEOVarunMohan,co−founderDouglasChen,andstaff,endingOpenAIacquisitiontalks);Gartnerpressrelease(August26,2025—40206,000 in 2025, up ~50,000YoY);EamesAI2026compensationdata(SanFranciscoAIagentengineer 252,000 median total compensation).
Comments are closed