When AI Agents Became Real: February's Quiet Revolution

February 2026 wasn't just another month of incremental AI progress. According to people actually building with these systems, it was the month the paradigm flipped—when autonomous agents stopped being a promising demo and became the actual way work gets done.

The shift is hardest to see if you're not in the weeds. But for developers and the increasingly large cohort of technically literate users who've been tracking model releases, the last two months represent something more fundamental than better benchmarks. As former OpenAI founder Andrej Karpathy put it: "It's hard to communicate how much programming has changed due to AI in the last 2 months. Not gradually and over time in the progress as usual way, but specifically this last December."

His diagnosis: "Coding agents basically didn't work before December and they basically do now." The models have enough coherence and tenacity to actually complete complex, multi-step tasks. Programming, he argues, "is becoming unrecognizable. The era where you type code into an editor is done."

Instead, developers are spinning up AI agents, giving them goals in natural language, and managing their work. The prize isn't the individual agent anymore—it's orchestration. How many can you run simultaneously in a way that produces something real?

OpenClaw and the Democratization of Agents

This abstract shift got a concrete manifestation in February: OpenClaw, the tool that went through three names in rapid succession (Claudebot, Moltbot, OpenClaw) as its relationship with Anthropic evolved. OpenClaw does what previous generations of AI assistants promised but couldn't deliver—it gives powerful language models actual access to your computer and lets them autonomously execute on complex tasks.

The homepage pitch is modest: clean your inbox, manage your calendar, check you in for flights. The reality is much wilder. Users immediately pushed past personal assistant territory into building autonomous agent teams—developers, researchers, project managers operating semi-independently.

What's genuinely interesting is who's using this stuff. OpenClaw is emphatically not beginner-friendly. It requires technical sophistication, tolerance for frustration, and willingness to debug opaque failures. Yet nearly 5,500 people are currently working through "Claw Camp," a self-directed program for building agent systems. These aren't all engineers.

CNBC's Deirdre Bosa tried building her own version of Monday.com using Claude Co-work, expecting it wouldn't work but figuring it would demonstrate current limitations. An hour later: "I literally have my own monday.com that's plugged into my calendar and Gmail and surface to kids bday that was not anywhere on my radar and I need to get a gift for."

Solopreneur Ben Sarah built Pulsia—an AI system for running autonomous AI companies—and reached $1.25 million in annual recurring revenue within weeks. You feed it a business idea (or ask it to generate one), and it handles everything from GitHub to Meta ads.

The labs are racing to catch up. OpenAI released the Codex app and hired OpenClaw's creator. Anthropic added remote control and scheduled tasks. Perplexity announced Perplexity Computer. Microsoft's Satya Nadella is reportedly using OpenClaw and encouraging his team to explore it. This "clawification" of AI—the shift toward actual agentic systems—is proliferating fast.

Wall Street's Panic

The second group that woke up in February was Wall Street, and their realization took a different form: stock selloffs labeled the "SaaS apocalypse."

The pattern was almost comically mechanical. Anthropic would announce a new Claude Code or Co-work plugin. Stocks in tangentially related sectors would crater. Gaming publishers dropped when Google demoed Genie 3's ability to generate immersive worlds. Productivity software, legal tech, finance—each took hits as agents demonstrated capabilities in their domains.

IBM had its worst single-day drop in 25 years because Anthropic blogged about its cobalt tool, which had been announced months earlier. By February 10th, both Bloomberg and the Wall Street Journal were writing about Wall Street's "hot new trade": dumping stocks in AI's crosshairs.

This created the perfect environment for Catrini Research's viral "2028 Global Intelligence Crisis" report, which outlined a theoretical doom loop leading to economic catastrophe. When Block announced 4,000 layoffs—40% of its workforce—many pointed to it as evidence of exactly the white-collar carnage Catrini predicted. (Others suggested it might be the year's most aggressive case of AI-washing.)

The market psychology matters here. Wall Street isn't spooked by infrastructure spending or hype cycles anymore. They're spooked because the tools might actually work—and quickly enough to destroy business models before new ones stabilize.

Washington's Messy Entrance

The third awakening was Washington's, and it was messier than anyone wanted.

The conflict centers on Anthropic's negotiations with the White House over AI use by the Department of Defense. Anthropic wanted specific redlines: no autonomous weapons, no domestic mass surveillance. The administration wanted a broader standard: any lawful uses.

But the disagreement isn't really about those specific applications. It's about who gets to decide how AI is used—and it's the opening salvo in what will be a much longer power struggle.

Senator Tom Tillis captured the dysfunction: "Why in the hell are we having this discussion in public? Why isn't this occurring in a boardroom or in the secretary's office? I mean, this is sophomoric."

It got less sophisticated from there. President Trump and Defense Secretary Pete Hegseth announced the government wouldn't work with Anthropic—and designated them a supply chain risk, potentially forcing other government contractors to drop relationships with the company.

The specifics of this particular fight matter less than what it represents: the inevitable collision between AI companies that grew up in Silicon Valley's move-fast culture and a government apparatus that operates on entirely different timelines and incentive structures. February was when that collision became unavoidable.

What Changed

Three groups experienced different versions of the same realization. AI insiders shifted their entire workflow around agent orchestration. White-collar workers discovered they could build complex systems without coding. Wall Street recognized the disruption was imminent and material. Washington understood the technology could no longer be managed through informal back-channels.

The model releases themselves—Opus 4.6, Gemini 3.1 Pro, Seed Dance 2.0, whispers of GPT-5.4—matter primarily as fuel for this shift. When benchmarking tools like Meter's long-horizon task study can't keep up with capability improvements, you're in genuinely uncharted territory.

Karpathy's framing holds: "This is nowhere near business's usual time in software." February wasn't about incremental progress. It was about a fundamental change in what's possible, who can access it, and what the second-order effects look like when those capabilities escape the lab.

The question for March isn't whether the pace continues—it's whether the systems we use to make sense of this change can keep up.

—Dev Kapoor