Anders Hejlsberg Says Developers Will Soon Supervise AI Agents, Not Write Code

Anders Hejlsberg, the legendary architect behind TypeScript, C#, and Turbo Pascal, has a stark prediction for the software industry: developers will soon spend more time supervising AI agents than writing code themselves. "The AI is doing the work, and you're supervising," Hejlsberg said in a recent GitHub video, describing a fundamental shift from AI as coding assistant to AI as autonomous agent.

The distinction matters. For the past two years, tools like GitHub Copilot have acted as glorified autocomplete—helpful, but still waiting for a human to type the next line. Now, according to industry analysis from nexos.ai and emerging platforms from Google and GitHub, we're entering a new phase where agents plan entire features, write tests, debug failures, and push code with minimal human intervention. The developer's role is morphing from craftsman to conductor.

Google made this vision concrete in late 2025 with Antigravity, an AI agent development platform that embeds autonomous agents directly into the coding environment. According to nexos.ai's early 2026 assessment, Antigravity transforms the development environment into "mission control for multi-agent workflows." Engineers define high-level outcomes—"build a user authentication flow" or "fix the checkout bug"—and the platform's agents break those goals into actionable steps, executing across the editor, terminal, and browser. Every task produces what Google calls "Artifacts": structured outputs like plans, test logs, and screenshots that give teams visibility into what the AI actually did.

This isn't theoretical. Devin AI, another tool highlighted by nexos.ai, already functions as "the closest thing we have today to a junior engineer who never sleeps." It understands entire repositories, plans development tasks, writes and debugs code, runs tests, and navigates live development environments. The architecture combines long-context language models with tool calling, sandboxed code execution, and task planning—making it a practical, programmable teammate within real-world workflows. The caveat: it requires well-scoped objectives and isn't safe for unsupervised production code merges. But for bug fixes, scaffolding, and unit testing, it already sets the bar.

The infrastructure layer matters as much as the agents themselves. Hejlsberg discussed the Model Context Protocol (MCP), a new standard designed to connect various development tools, environments, and workflows so AI-driven systems can access and coordinate between them efficiently. Without MCP or similar protocols, agents remain isolated—brilliant at one task but unable to orchestrate across the messy reality of modern software stacks. With it, an agent can pull context from your CRM, update your documentation, spin up a test environment, and notify your team in Slack, all as part of a single workflow.

German consultancy Ventum Consulting is already teaching enterprises how to operationalize this shift through what it calls "Agentic Coding" workshops. According to Ventum's March 2026 materials, companies are learning to go "from idea to specification to finished app in 1 day" using GitHub Copilot, Claude, and spec-driven frameworks like SpecKit. The workshop teaches developers to write precise natural language specifications that agents convert directly into code, then systematically review and refine through iterative loops. The emphasis is on structured specs, tests, guardrails, and clean repository templates—not ad hoc prompting.

This spec-driven approach addresses one of the biggest risks in agentic coding: garbage in, garbage out. As nexos.ai notes, "high-quality context is crucial for AI results." Developers must learn to provide agents with the exact context they need—no more, no less—to avoid hallucinations, token bloat, and wasted compute. Ventum's workshops drill teams on "the 'Every Token Counts' principle," teaching them to structure code, documentation, and specifications so agents can consume them efficiently. The result: faster development, higher quality, and lower AI costs.

The business implications are profound. nexos.ai, which offers an all-in-one AI platform for teams of 50-500 employees, reports that companies are now deploying no-code agents across marketing, sales, operations, and people teams—not just engineering. These agents book meetings, summarize documents, fill forms, update CRMs, and run multi-step workflows with minimal oversight. The platform provides role-based access control, audit trails, and full observability of agent behavior, addressing the governance and security concerns that have slowed enterprise AI adoption.

But Hejlsberg's most provocative claim is that traditional IDEs—the integrated development environments that have defined programming for decades—may soon become obsolete. If agents are doing the heavy lifting, why do developers need complex interfaces designed for line-by-line editing? Future development environments, he suggests, may look fundamentally different: less about syntax highlighting and more about task orchestration, agent monitoring, and quality control dashboards.

We're not there yet. Antigravity is still in public preview and requires human oversight for production work. Devin AI is effective but needs well-scoped tasks. Even the most advanced agents struggle with ambiguous requirements or legacy codebases with poor documentation. And as nexos.ai warns, "AI agent performance characteristics change frequently"—what works today may be obsolete in six months.

Still, the trajectory is clear. The question isn't whether AI agents will write most code, but when—and whether developers will adapt fast enough to stay relevant in a world where their primary skill is knowing what to ask the machines to build. Hejlsberg, who has spent four decades shaping how developers work, seems to think the answer is soon. The best developers will be those who learn to supervise, not just code.