Relixir Is Building AI Employees That Improve Themselves

The Macro: AI Agents Are the New Hires

The AI agent gold rush is real and getting louder. Every week there’s another startup claiming their agent can replace a full-time employee. Most of them are wrappers around language model APIs with some prompt engineering and a nice dashboard. The actual hard problem, building an agent that reliably performs a complex job function without constant human oversight, remains mostly unsolved.

The companies making real progress tend to pick a narrow job function and go deep. Klarna replaced 700 customer service agents with AI. Harvey focuses on legal work. Cognition’s Devin targets software engineering. The pattern is clear: horizontal “do everything” agents fail. Vertical “do this one thing really well” agents ship.

But there’s a second problem lurking beneath the first. Even agents that work well on day one tend to plateau. They don’t learn from their mistakes. They don’t adapt to changing conditions. They don’t get better at their jobs the way a human employee would over six months. The feedback loop is manual: a human reviews the output, tweaks the prompts, adjusts the workflow. That’s fine for a proof of concept. It doesn’t scale.

This is where the concept of recursive self-improvement enters the conversation, and where things get both interesting and speculative.

The Micro: Berkeley Dropouts, Self-Improving Agents, 400+ Customers

Sean Dorje and Dennis Zax dropped out of Berkeley to build Relixir. They came through YC’s Spring 2025 batch and have backing from DEEPCORE (SoftBank’s deep tech fund) and the founder of Turing. The team is four people, based in San Francisco.

The core technology is what they call “autonomous agent runtimes with recursive self-improvement.” In practice, this means their agents have built-in mechanisms to evaluate their own performance, identify where they’re falling short, and modify their behavior without a human stepping in to retune them. That’s the RSI part. It’s a bold claim, and the degree to which it actually works versus being an aspirational label is the key question.

Their first commercial agent is Rex, which focuses on GEO, or Generative Engine Optimization. If you haven’t encountered this term yet, you will. GEO is the practice of getting brands recommended by AI search engines like ChatGPT, Perplexity, and Claude. Traditional SEO optimizes for Google’s link-based ranking. GEO optimizes for how large language models talk about and recommend products. It’s a new field, barely a year old, and there’s no established playbook.

Rex handles GEO autonomously for brands. The customer list is real and impressive for a company this young: Rippling, Airwallex, HackerRank, and 400+ others. That’s not a handful of beta testers. That’s a real customer base paying for a product that works well enough to keep using.

The broader vision goes beyond marketing. Relixir positions itself as building agent runtimes that can automate entire organizational functions, with the eventual goal of enabling “zero-human company operations.” The website states plainly that “in the next economy, large organizations will downsize headcount.” Whether you find that exciting or unsettling probably depends on which side of the automation line you sit on.

They also offer a product called Naive (at usenaive.ai), which appears to be a more general company agent runtime. The relationship between Rex and Naive in their product strategy isn’t entirely clear from the outside, but it suggests they’re building a platform, not just a point solution.

The Verdict

I think Relixir is one of the more credible entries in the AI agent space, primarily because of the traction. Four hundred customers including names like Rippling and Airwallex is real validation, not investor marketing. The GEO focus is smart because it’s a new category with no dominant player, and the need is genuine as AI search engines eat into traditional search traffic.

The recursive self-improvement angle is harder to evaluate from the outside. Every AI company claims their product “learns and improves.” Most of them mean “we retrain the model quarterly.” If Relixir’s agents genuinely improve their performance autonomously in real-time, that’s a meaningful technical moat. If it’s more incremental than that, they’re still a strong GEO tool with good traction, just not the paradigm shift they’re marketing.

The team is young and small, but that hasn’t stopped them from landing enterprise customers. The risk is focus. Building a GEO agent, a general company agent runtime, and a recursive self-improvement platform simultaneously with four people is a lot of surface area. The companies that win in agents tend to be the ones that resist the urge to go horizontal too early.

I’d bet on the GEO product specifically. The market is new, the timing is right, and the customer list suggests they’ve found something that works. The broader “automate all the humans” vision is further out and carries more execution risk. But if Rex keeps delivering results and the self-improvement claims hold up, this is a company that could define how the GEO category develops.