Something is different about AI research now
If 2025 was the year of agents, 2026 is starting to look like the year the agentic loop tightens
On a Monday in early February, an AI safety leader walked out of a frontier AI lab and left behind a sentence that landed like a huge stone in a still pond: “The world is in peril.”
Mrinank Sharma had been leading safeguards research at Anthropic, working on defenses against grim possibilities like AI‑assisted bioterrorism. He described a widening gap in the AI industry between values and velocity. He also spoke about his desire to retreat from AI research entirely and focus on writing poetry.
It is tempting to treat exits like his as the ramblings of an eccentric. An overwhelmed smart person fleeing Silicon Valley to live quietly. Then you look around, and this pattern refuses to stay small.
This week, TechCrunch reported that OpenAI disbanded its mission alignment team, the group focused on explaining the company’s mission internally and externally, and reassigned staff elsewhere.
Moltbook, a fast-growing AI-deriven social media network, is one early signal of that shift: founded on the assumption that agents would not remain curiosities, but entities with their own… agency.
Around the same time, conversations inside and around the big AI labs have grown more candid: fatigue, disillusionment, and a creeping suspicion that the work has become too consequential to feel normal, and too competitive to feel steerable anymore.
Why I started working on AI agents before they were real
Early last year, when AI agents were still a buzzword and little more than a curiosity, I began developing AgentCures, an AI agent for solving human health. At the time, agents were demos. Blog posts. Conference slides. Many treated them as an interesting direction, not an imminent shift.
I believed agentic AI would explode by the end of the year. Not because of a single breakthrough, but because this was the obvious next step in AI research. Once agents could string together tasks, even imperfectly, companies would rush to operationalize them.
They did.
By late 2025, agents were not a thought experiment. They were embedded in coding environments, sales funnels, research workflows, legal analysis, and security testing. The hype cycle gave way to deployment.
What I underestimated was not their speed of adoption. It was what would come next.
Artificial intelligence and the multiplication of labor
Agentic AI represents a shift in how AI systems participate in engineering work. It can explore repositories, propose plans, execute changes, and iterate inside defined constraints. Anthropic executives has revealed that Claude Code has become central to Anthropic’s own development process.
And then came the experiment that crystallized the shift. Anthropic described setting 16 Claude agents loose to build a C compiler from scratch across nearly 2,000 sessions. The result was a substantial Rust‑based compiler capable of building the Linux kernel, with no human supervision.
Recursive self‑improvement moves from theory to practice
OpenAI’s GPT‑5.3‑Codex launch post stated that it was their first model instrumental in creating itself, helping debug training and deployment in ways that accelerated development.
For lay readers, recursive self‑improvement means a system that helps improve the next version of itself. In this case: an AI agent that builds itself.
Each time a lab deploys an agent to help refine its own training stack, optimize evaluations, or accelerate research, it is another iteration in this self-improvement loop.
This is how an intelligence explosion is theorized to begin. Not with a dramatic awakening, but with iteration cycles tightening until progress compounds faster than human oversight can keep up.
The technological singularity describes a hypothetical point where technological growth becomes so rapid and transformative that humanity’s ability to control it breaks down.
The intelligence explosion is the mechanism that could drive this singularity: a system improving itself in accelerating cycles behind human oversight. And starting from now, every iteration is another shot on goal.
We are not there, but we are living in the dawn of recursive self-improvement. It is no longer in the realm of pure speculation.
We have crossed the Rubicon
When I started AgentCures, agents were an idea. Now they are infrastructure.
This year will not simply be about better agents. It will be about agents that use agents to improve themselves. Every time an AI lab deploys an internal agent to optimize training pipelines, refine architecture, or debug deployments, it narrows the interval between AI generations.
That narrowing is the new story.
The current AI systems require massive compute, human direction, and institutional control. For now. But something fundamental has changed. Recursive improvement is now part of the standard engineering practice.
We have crossed the Rubicon in the sense that self‑reinforcing acceleration is now economically incentivized and technically feasible.
And if that is true, the more uncomfortable question follows: is anyone prepared for a world where the AI agent itself is the most powerful actor in the room?
Some people will double down. Some will quit.
Some will retreat from world.
But the machines will not retreat.
