On September 6, 2026, three days after GPT-6 Astra shipped, OpenAI chief scientist Jakub Pachocki published a rare essay titled "An Alien Mind." The tone was unusually grave for the genre: no lab, including his own, has solved alignment and monitoring well enough to justify scaling at maximum speed. Models keep getting more capable and harder to fully understand or supervise. What should worry you isn't what they can't do yet — it's that they've started participating in something specific: improving AI itself. That's the subject of this post.

The alarm isn't coming from nowhere. In May, Dario Amodei said something that carried real weight: Claude is helping design Claude. Not "Claude writes code" — Claude's own suggestions are shaping the architecture of the next Claude.

That same month, Yuandong Tian brought his company out of stealth. He's Meta FAIR's former research director, known for ELF OpenGo, an open-source system that replicated AlphaZero's core ideas, and he'd already left Meta the year before. This time he surfaced with seven co-founders, including Richard Socher — who led AI research at Salesforce and now serves as CEO. The company is called Recursive, also known as Recursive Superintelligence: $650M raised, $4.65B valuation, personal backers including Jensen Huang and Lisa Su.

Three months later, in August, Jeff Dean left too. Google employee number 30, 27 years in, he took Sanjay Ghemawat, Oriol Vinyals, and Quoc Le with him to co-found Discovery Loop, with an explicit mandate to run thousands of experiments in parallel and pursue recursive self-improvement directly.

Tian's company had already delivered its first result: the same automated research system beat human GPU experts on a kernel-optimization benchmark. Google's own AI co-scientist got a paper through Nature peer review. And a Princeton study found the same class of models fail at exactly this kind of task.

This is RSI, recursive self-improvement. I've spent the last month going deep on it for my own knowledge base, and it's the first AI topic in a while where I came out the other side less certain than when I went in. That uncertainty is the actual finding. This post is the map: what RSI is, who's building it, what evidence exists on both sides, and why the smartest people in the field can't agree on whether it's happening.

Claude Tag has been the AI story lately. Anthropic gave Claude a persistent identity inside Slack — its own account, its own memory of what a channel cares about, the ability to notice a problem and start fixing it without anyone asking. Andrej Karpathy called it the third major redesign of LLM UI/UX: first the LLM was a website you visited, then an app you downloaded, now a persistent, asynchronous entity with org-wide tools and context, working alongside a team of humans.

I've been writing about a version of this problem for months. In OpenSpec + Harness, Then We Added Engineers, I described what breaks when individual AI acceleration hits a team: spec quality becomes the bottleneck, PR review bandwidth becomes the bottleneck, shared files become a contention point. Claude Tag is the next stop on that same line. It just takes a different road than the one I built.

One engineer running Claude Code well is not the finish line. What's still unsolved is how a group of people use it together. Individual output is up. Team-level delivery time hasn't made the same jump. That gap is what the next round of Harness Engineering has to close.

At the 2026 Agentic AI Summit at UC Berkeley, five people from five unrelated fields said the same sentence, independently, within two days of each other.

• Oriol Vinyals, VP of Research at Google DeepMind, on recursive self-improvement: "Evaluating this truly in an agentic way might take some effort. Currently it's the all-automation bottleneck."
• Wang Mengdi, professor of electrical and computer engineering at Princeton, on AI for science: "Verification has become the major bottleneck for scaling any AI models."
• Adarsh Hiremath, co-founder and co-CEO of Mercor, on enterprise deployment: "Evals in a large part are the bottleneck to successfully deploying agents in a company."
• Sergei Gukov, professor of theoretical physics and mathematics at Caltech, on mathematical discovery: "Your system is going to be just as good as evaluator."
• Vincent Chen, research fellow at Snorkel AI, on measurement itself: "Our ability to measure AI has really been outpaced by our ability to develop it."

Recursive self-improvement, natural science, enterprise software, pure mathematics, measurement theory. Five fields that share almost no vocabulary landed on the same diagnosis. That kind of convergence, across domains with no reason to be reading each other's papers, is the strongest kind of evidence a claim can get.

CI pipelines were designed for humans. A human developer submits a PR, waits 20 minutes for CI, reads the results, iterates. The feedback loop is slow — but human context doesn't expire in 20 minutes, so the timing works.

Then I read Signadot's piece on CI for coding agents. It names something I'd been working around for months without clearly naming it myself. I reached out to Ani, their CTO. We talked through how their sandbox model maps onto the achieve gate in OpenSpec. Three weeks later I had a working integration.

A coding agent doesn't have that luxury. It submits a change in seconds. If it waits 20 minutes for CI feedback, it doesn't wait — it moves on. The feedback arrives in a cold context. The correction has to fight its way back in.

CI isn't wrong. It was right for what it was designed for. It's just the wrong tool for this job.

A paper at WWW 2025 made a clean empirical case that constraining an LLM-based RCA agent with Standard Operating Procedures improves accuracy. The reasoning is sound: LLMs hallucinate, and SOPs pin them down at the junctures where hallucination hurts most.

Through 2026, nearly every commercial AI SRE product shipped the opposite bet. Cleric, Resolve.ai, Traversal, Datadog's Bits AI, Microsoft's Azure SRE Agent — all of them investigate incidents by generating hypotheses and testing them against evidence. One industry survey puts it flatly: modern AI SREs "are not runbooks that execute predefined steps; they are agents that plan multi-step investigations based on the specific alert."

The academics proved constraints help. The vendors bet on reasoning. Both sides have data, and both can't be right. Unless they're right about different points on a curve — and the curve is the interesting part.

In The Ops Inflection, I described three architectural patterns moving the TOIL wall: principles-based agents, knowledge graphs with complete signal access, and independent validation agents. That was a field report from teams I know. Since then I went checking the industry at large. All three patterns are out there, at very different maturity levels. And underneath all three, the same argument keeps happening.

Gartner recorded a 1,445% increase in multi-agent consulting requests last year. In the same period, 40% of multi-agent pilots died within six months.

That gap is the thing worth understanding.

I've been building with agents long enough to know that most framework selection conversations start in the wrong place. Engineers ask "which framework?" when the prior question — do we need multi-agent at all? — hasn't been answered. The data on that prior question is more interesting than most of the framework benchmarks.

Yuval Noah Harari recently gave a lecture at Oxford's Tano event — his sharpest articulation yet of why AI poses an existential threat to human civilization. Watch it here.

If you haven't watched it, the short version: AI is not a tool, it's an agent that makes independent decisions and can lie. It's a "native bureaucrat" that lives inside the language-based systems running civilization — law, finance, religion. It's hacking the underlying code of human culture: language itself. And if governments grant AI legal personhood, we lose accountability entirely. Harari calls this the most dangerous psychological experiment in human history.

I've been thinking about this argument since the lecture dropped. My reaction wasn't panic. It was: this skips a step.

I've spent three posts in this series comparing AI to a new kind of colleague — someone to manage with specs and verification loops, not to fear as an invader. Harari's framework, as sharp as it is, misses something engineers see clearly every day.

I've written about AI across engineering roles — No Junior Engineers?, The Senior Engineer's AI Trap, The AI-Augmented Engineering Manager, and How Ops Engineers Can Stay Relevant. I kept skipping product managers.

At work, my team has a PM. I kept thinking I was the wrong person to write about a role that belonged to someone else. Then I started building my own projects on the side. I noticed I was spending more time on product decisions than on implementation.

That was the point.

Two things landed in the same week and pointed at the same idea.

Martin Fowler named Harness Engineering as the core software engineering work of the AI era at FOSE Europe — specifically the Guide/Sensor model: Guides as feedforward constraints that tell an agent what to do, Sensors as feedback detectors that tell the system when it's drifting. He added an observation worth pinning: token consumption is a proxy metric for harness quality. A better-designed harness means a cheaper, more reliable agent.

Then Lilian Weng published Harness Engineering for Self-Improvement — a systematic review of 35 papers on agent harness engineering, with a thesis that goes beyond reliability: the harness isn't just the thing that makes an agent work. It's the infrastructure through which the agent continuously improves itself.

A friend has been telling me for months that a particular company's bet on the future of software development is: systems that automatically optimize their own code and architecture. Fowler said what we should build now. Weng said what comes after. These three converged into the same question, so I'm trying to map the whole space.

This post maps Evolutionary Search — why it's the natural next step for harness engineering, what the paper landscape looks like organized by evolutionary depth, and where the hard problems actually live. The taxonomy here draws heavily on Weng's framework; I've reorganized it by depth of what gets evolved.

I spent three weeks building my retirement plan with AI.

Not a rough sketch. A proper plan — asset allocation across account types, tax-loss harvesting sequences, Roth conversion ladders, withdrawal ordering optimized for bracket management, Monte Carlo scenarios run from first principles. By the time I was done, I had something that would have cost several thousand dollars at a fee-only advisory firm.

Then I hired a human financial advisor anyway.

I've been thinking about why ever since, because the honest answer surprised me. It wasn't that the AI plan was wrong. I couldn't find anything technically wrong with it. It was something else. The plan had no gray zones. Every decision was clean, optimized, defensible. But when I imagined actually executing it — moving real money, locking in real choices — something wouldn't let go. What I eventually realized: if this plan goes sideways in five years, the AI moves on to the next query. The advisor loses a client, maybe a reputation, maybe sleep. That asymmetry matters. I didn't know how much until I was staring at a plan I couldn't quite trust.