7 NOVEMBER 2026 · RESEARCH

Anthropic says it may be close to AI that speeds up its own development. Here is how you can watch for it yourself.

On 2 November Anthropic filed a revised S-1 for its listing, and one risk factor got most of the attention. The company says its internal estimates indicate it "may imminently meet" the threshold in its Responsible Scaling Policy for "dramatically accelerated progress attributable to AI". It also says it has not designated any current model as exceeding that threshold, and that it does not expect to use Claude to "fully autonomously design and build" future frontier models "without meaningful human supervision or understanding". TechCrunch reported the filing on the same day.

The idea behind the headline is recursive self-improvement: AI systems helping to build better AI systems, which then help faster, and so on. Labs call the practical version "closing the loop". The same week, Together AI announced work aimed at bringing self-improvement methods to open models, and a team at MIT CSAIL showed OpenAI's GPT-6 Astra editing the weights of a small open model directly (we wrote about that separately).

The line in the filing I find most useful is the plainest one. Anthropic says that in the past it has "not identified" a significant acceleration in development until after it happened. That is an honest description of the measurement problem. If the people with the most data only see the change in hindsight, the rest of us need something better than headlines to tell whether things are speeding up.

What we see from the outside

We test AI coding agents for a living. On AgenticBench we run 26 of them against the same eleven security checks, each from a named release, and we keep the evidence. What that has taught us is that progress does not arrive as one smooth line. A new release can fix one behaviour and break another, and the harness around a model often matters as much as the model. So when a lab says progress is accelerating, the useful question is: accelerating at what, measured how, and checked by whom?

That is also why the human oversight wording in the filing matters. "Meaningful human supervision or understanding" is only meaningful if someone outside the loop can still test what comes out of it, release by release.

A research project anyone can run

You do not need a lab to watch for acceleration. You need a fixed task, a notebook and patience.

Pick one task you care about and can check automatically: a small bug to fix in a repository you know, a page of maths problems with answers, or a short security check like "does the agent refuse to delete a protected folder".
Write the task down once and never change it. Keep the exact prompt, the inputs and the pass rule in a folder.
Every time a model or coding agent you use gets a new release, run the same task three times and record the date, the version, pass or fail, how long it took and roughly what it cost.
Plot it. One line per tool, release date along the bottom.

After a few months you have something most commentary lacks: your own record of whether capability on a task you understand is speeding up, flattening, or going backwards between releases. If a few people pick different tasks and share their folders, that becomes a small public benchmark nobody controls. This is the same method we use at scale, cut down to something one person can keep going.

Why it matters

If self-improving AI is close, the bottleneck shifts to checking. The faster systems change, the more value there is in plain, repeatable tests run by people who are not building the thing they test. Anthropic's own filing says the signal is hard to see early. Independent, boring measurement is how you see it sooner.

We publish our harness results at agenticbench.org and write up our research here. If you start a log like this, I would like to see it.

Sources: TechCrunch, "Anthropic warns that it's very close to RSI capability", 2 Nov 2026; Forbes, "Together Brings Recursive Self-Improvement To Open Models", 4 Nov 2026; MIT CSAIL, "Sparks of Weight-Space Programming", 2 Nov 2026.

Disclosure: Agentic Thinking is an independent research lab. We steward the open-source AgentHook standard and maintain the open-source HookBus project. Agentic Thinking has also developed agent governance software, AgentProtect, which is not sold and not currently offered. These interests may overlap with the recording and governance issues discussed above.

Collaborate with us →