The market has delivered a clear verdict on how AI should reshape software development: assistance beats replacement. Cursor, an AI-powered code editor built by Anysphere, reached roughly $4 billion in annualized revenue in under four years, with approximately $2.6 billion coming from enterprise customers. By comparison, Cognition's autonomous AI engineer Devin generated $492 million in annual revenue, making Cursor's assistance model roughly eight times more profitable. In June 2026, SpaceX acquired Anysphere in an all-stock deal valued at $60 billion, the largest acquisition of a venture-backed startup ever recorded, underscoring the strategic value investors now place on tools that keep humans in control.

Why Are Developers Choosing Assistance Over Autonomy?

The revenue gap reflects how software organizations actually adopt new technology. An assistant like Cursor integrates directly into a developer's existing workflow and delivers value on day one without requiring the organization to restructure how it builds software or whom it trusts with critical decisions. A developer using Cursor becomes more productive immediately, and the value compounds without anyone needing to decide that a machine should be accountable for shipped features. Autonomous agents, by contrast, ask for something far heavier: a fundamental reorganization of responsibility in which the human moves from author to overseer, and the agent is trusted to complete work unsupervised. That represents a higher adoption cost, a steeper trust threshold, and a slower path to revenue.

The quality of the code matters enormously in this equation. Research from Apiiro found that AI-generated code introduced more than 10,000 new security findings per month across the repositories studied, a tenfold increase in six months, with privilege-escalation paths up 322 percent and architectural design flaws up 153 percent. An analysis by DryRun Security found that coding agents introduce security vulnerabilities across nearly every category of application they build, often omitting security components entirely or constructing flawed authentication logic. Independent benchmark work has found that a substantial share of AI-generated code, by some measures 40 to 62 percent, contains security vulnerabilities or design flaws.

These defects do not disappear when an autonomous agent ships them. They move downstream into production, where someone else absorbs the cost. A Stanford analysis found that reworking AI-generated code consumes 15 to 25 percentage points of the 30 to 40 percent productivity gains the tools nominally deliver, which means a meaningful fraction of the promised acceleration is spent cleaning up afterward. This is exactly where the assistance model holds its structural advantage. Keeping a skilled human in control of the work means the review happens continuously, at the moment of authorship, rather than after a finished artifact has already been committed. The assistant catches the flaw before it ships. The autonomous engineer discovers it in the incident log.

What Are the Long-Term Risks of Removing Humans From the Loop?

Beyond immediate code quality, the longest-horizon argument for assistance concerns the talent pipeline. The tasks an autonomous engineer is most capable of taking over, namely boilerplate code, standard patterns, routine debugging, and unit tests, are the same tasks through which junior engineers have historically built judgment and expertise. By automating these foundational learning experiences, the industry risks eroding the next generation of skilled developers who understand how to reason about code architecture and security.

Autonomous agents are gaining real enterprise traction despite these concerns. Cognition reports enterprise customers including Goldman Sachs, Citi, Mercedes-Benz, Dell, Santander, and units of the U.S. Army and Navy, with enterprise usage growing more than tenfold since January 2026. Mercedes-Benz reported compressing a legacy modernization project from eight months to eight days, and Cognition disclosed that roughly 89 percent of the code committed by its own engineers is now written by Devin. These hyper-compressed timelines and coding feedback loops carry significant risks, and Cognition's foothold within major enterprises and government means that these risks could carry profound consequences for consumers and the American citizenry.

How to Evaluate AI Coding Tools for Your Team

• Adoption Friction: Assess whether the tool integrates into your existing developer workflow or requires restructuring how your team builds software and manages responsibility for shipped code.
• Code Quality and Review: Evaluate whether the tool keeps humans in the review loop at the moment of authorship or ships finished artifacts that require downstream remediation and security audits.
• Talent Development: Consider whether the tool automates routine tasks that junior engineers use to build foundational judgment, or whether it preserves learning opportunities that develop the next generation of skilled developers.
• Security Posture: Review independent benchmarks on the rate of security vulnerabilities and design flaws in the tool's output, and understand what percentage of promised productivity gains are consumed by reworking defective code.

The market's preference for assistance over autonomy is not sentimental. It reflects the economic reality that a tool generating $4 billion in revenue from customers who trust it to make them faster is more valuable than a tool generating $492 million from customers who are still learning to trust it with unsupervised work. Cursor's acquisition price signals that investors believe the assistance model has already proven its business case, while the autonomous model is still being priced on what it might become.

The debate between assistance and replacement will likely continue as autonomous agents improve and prove their reliability in more domains. But for now, the evidence suggests that the most valuable AI in software development is not the one that removes the human from the loop, but the one that makes the human faster while keeping them in control.