Orbit | From Vibe Coding to Agentic Engineering: What the $285B SaaSpocalypse Means for How We Build Software

On February 3, 2026, $285 billion evaporated from software stocks in a single day. Bloomberg coined it the "SaaSpocalypse."

The catalyst: Anthropic released Cowork plugins for legal, finance, and data marketing — AI agents capable of replacing entire SaaS workflows. Goldman Sachs' basket of US software stocks sank 6% in one session. Thomson Reuters plunged 16%. LegalZoom dropped 20%. The ripple spread to financial stocks — LPL Financial fell 8.3%, Charles Schwab 7.4%.

The same week, Andrej Karpathy — the person who gave us "vibe coding" exactly one year earlier — declared it passé and introduced its successor: agentic engineering.

These aren't separate stories. They're two sides of the same shift. The tools we build software with are no longer assistants. They're autonomous agents. And the entire industry is repricing around that reality.

Vibe Coding's Birthday Party Turned Funeral

February 2025: Karpathy posted on X, describing a workflow where you give in to the vibes and forget the code even exists. By year's end, "vibe coding" was the Collins Dictionary Word of the Year. Stack Overflow's 2025 developer survey found 65% of developers using AI coding tools weekly. AI was writing 30% of Microsoft's code and over 25% of Google's.

But the cracks were showing.

METR's study revealed that experienced open-source developers were actually 19% slower with AI assistance — despite believing they were 20% faster. CodeRabbit's analysis found 1.7x more major issues in AI co-authored code. The gap between perceived benefit and actual benefit was significant.

Then came the academic reckoning. Researchers from Central European University and the Kiel Institute published a paper arguing that AI-mediated development erodes the open-source ecosystem. Their evidence: Tailwind CSS documentation traffic dropped 40% since early 2023 despite rising usage. Stack Overflow activity fell roughly 25% within six months of ChatGPT's launch. Tailwind CSS revenue declined close to 80%, forcing layoffs. The paper hit the front page of Hacker News with over 285 comments.

The core problem: vibe coding decoupled usage from understanding. Developers consumed open-source libraries through AI intermediaries without ever visiting documentation, filing bug reports, or contributing back. When humans stop engaging with the code, the ecosystem that produces it degrades.

Karpathy himself acknowledged the limitation. Vibe coding, he noted on its anniversary, was always meant for throwaway weekend projects. It was never meant for production.

The Week Wall Street Priced In the Agent Era

The SaaSpocalypse wasn't a correction. It was a repricing.

When Anthropic launched Cowork, investors suddenly grasped a simple math problem: if one AI agent can do what ten humans with ten SaaS subscriptions did, the per-seat model collapses. Salesforce was down 26% year-to-date. ServiceNow down 28%. Intuit down 34%. ABC News reported it as the largest AI-displacement selloff in market history.

CNBC ran what might be the most effective product demo in financial journalism history: their reporters — non-developers — built a Monday.com clone in under an hour for $15 in compute costs. Monday.com has a $5 billion market cap.

Not everyone agreed with the panic. Jensen Huang called the selloff the most illogical thing in the world, arguing AI would enhance software companies, not replace them. Wedbush analysts called the SaaSpocalypse scenario far from reality.

The truth sits somewhere in between. AI agents won't replace every SaaS product tomorrow. But the direction is unambiguous — and $285 billion says the market knows it.

Enter Agentic Engineering

On February 4, Karpathy posted his new framework. His definition of agentic engineering: "agentic because you are not writing the code directly 99% of the time, you are orchestrating agents... engineering to emphasize that there is an art & science and expertise to it."

The distinction from vibe coding is precise. Vibe coding meant trusting the output, accepting AI suggestions without deep scrutiny, embracing the vibes. Agentic engineering means orchestrating agents with the same rigor a senior engineer brings to code review — oversight, quality gates, architectural decisions, specification discipline.

The infrastructure arrived the next day. Anthropic released Claude Opus 4.6 with agent teams — a feature that splits work across multiple coordinating AI agents working in parallel. The model set new benchmarks: 65.4% on Terminal-Bench 2.0 (the highest score recorded), 80.8% on SWE-bench Verified, an 83% improvement on ARC AGI 2 over its predecessor, and a 1-million-token context window in beta.

Agent teams is the technical infrastructure that makes agentic engineering real. Instead of one AI assistant helping one developer write code, multiple specialized agents coordinate — one exploring the codebase, another writing tests, another implementing features — with a human orchestrating the effort.

Anthropic CEO Dario Amodei pushed the timeline further, predicting that software engineering could become "completely obsolete" in 6-12 months. What he means is more nuanced than the headline suggests: the nature of the work changes, not that humans become irrelevant. Someone still needs to define what to build, evaluate whether it's correct, and make tradeoff decisions that AI can't.

The emerging skill set for agentic engineering looks nothing like traditional software development:

Prompt architecture — structuring complex instructions across multiple agent interactions
Agent orchestration — coordinating parallel agents and managing their outputs
Quality gates — defining verification criteria before agents execute
Specification writing — translating requirements into unambiguous agent instructions
Code review at scale — reviewing AI-generated output across larger codebases faster

What This Means for Developer Tools

The IDE is no longer a text editor with autocomplete. It's becoming an agent orchestration platform.

The market agrees. According to SaaStr, Cursor raised $2.3 billion at a $29.3 billion valuation. Replit did $240 million in revenue in 2025 — up from $2.8 million in 2024 — and is targeting $1 billion in 2026. Lovable crossed $300 million ARR. Combined, the top vibe coding startup valuations exceed $36 billion.

The creation bottleneck is broken. iOS app releases surged 60% year-over-year in December 2025 after years of flat growth — a step-function change aligned precisely with the rise of agentic coding tools.

But creation without governance produces exactly the problems the open-source paper describes. More software gets built, but less of it gets maintained, understood, or secured. The tools that win this era need to combine three things:

Agent orchestration — coordinate multiple AI agents on complex tasks
Human-readable workflows — make agent actions visible and reviewable
Quality enforcement — maintain standards as output velocity increases

The developer tools market is splitting. On one side: tools optimized for speed, for getting something working as fast as possible. On the other: tools optimized for sustainable development, where agents and humans collaborate with appropriate oversight.

The Case for Staying in the Loop

Amodei's prediction deserves examination against the evidence.

The METR study — experienced developers 19% slower with AI — remains the most uncomfortable data point for AI optimists. Bain's consulting research found real-world productivity gains from AI coding tools "unremarkable" compared to the hype. MIT Technology Review named generative coding a 2026 Breakthrough Technology while simultaneously publishing the research showing its limitations.

SWE-bench scores jumped from 33% in August 2024 to over 70% in 2025. But benchmarks measure task completion in controlled environments. Production software lives in uncontrolled ones — with legacy code, edge cases, shifting requirements, and users who do unexpected things.

The open-source crisis is the canary. When developers stop reading documentation, filing issues, and contributing patches — when AI mediates the entire relationship between developer and dependency — the ecosystem that every modern application sits on starts to erode. That's not a theoretical concern. It's happening now, measured in traffic drops and revenue collapses.

Agentic engineering's answer: you're still in the loop, but the loop is different. You're reviewing, not writing. You're architecting, not typing. You're orchestrating agents and verifying their output instead of producing code yourself.

This is harder than vibe coding, not easier. Reviewing code you didn't write requires deeper understanding than writing it yourself. Orchestrating multiple agents requires architectural thinking. Defining quality gates requires knowing what good looks like. The "engineering" in agentic engineering isn't decorative. It's the whole point.

The Shift Has Already Happened

The SaaSpocalypse wasn't really about AI replacing software. It was the market waking up to the fact that the relationship between humans and code has permanently changed.

Vibe coding showed what's possible when anyone can create software. Agentic engineering is the discipline that makes it sustainable. The tools, models, and market incentives are all pointing in the same direction: the developer's role is evolving from writer to orchestrator.

The question isn't whether this shift happens — $285 billion says it already has. The question is whether your tools are built for the old paradigm or the new one.

Orbit is building the development environment for the agentic engineering era — where AI agents, visual workflows, and code editing converge in a single workspace.

Join the waitlist →