When AI Builds AI

In February 2026, Claude led less than 1% of Anthropic’s measured AI research and development work. By August, it led 26%.

AI is increasingly writing the code, running the experiments and optimising the infrastructure behind its own development. That creates a powerful feedback loop: better models help build better models, while more efficient infrastructure makes the process cheaper.

The implications extend beyond developer productivity. If AI accelerates the work that produces the next generation of AI, what happens to the pace of improvement? Can oversight keep up? And where can startups build valuable businesses?

The feedback loop is taking shape

Those figures come from Anthropic’s R&D Automation Index, published in September. “Leads” means Claude could complete most of a task from a high-level prompt, with humans supervising. More than 90% of the measured work involved AI at least as a collaborator, and none was classified as fully autonomous. These are self-reported numbers, and Claude itself was used to build the task taxonomy and classify the work, a bias risk Anthropic acknowledges. Even so, the direction is clear: researchers are increasingly using today’s models to build tomorrow’s.

Together AI offers another concrete example. Its Together Compiler, announced on 4 November, uses agents to optimise the software that runs AI models, including low-level GPU kernels and how computation is distributed across hardware.

Inference workloads constantly change. New models arrive, prompts get longer and customers adopt different usage patterns. A configuration optimised for yesterday’s traffic may be inefficient tomorrow. Together’s agents search for better implementations, testing changes through a persistent engine that avoids repeatedly restarting the system.

The company reports up to 2.6 times higher token throughput than the best published open-source configurations on SemiAnalysis’s InferenceX benchmark, in less than 24 hours. It attributes much of that to running 252 times more experiments than agent systems that restart the engine for every test. These are vendor-reported results for particular workloads, but they illustrate a shift towards software that can be continuously optimised as conditions change.

One surprising finding in the Together Compiler research paper: giving agents an expert-curated knowledge base reduced their speedups by 25 to 31% after 12 hours, and naive agents told simply to make the system as fast as possible beat agents following expert-prescribed strategies. That does not make expertise redundant. It suggests its value may shift towards defining objectives, designing tests and deciding which changes are safe to deploy.

Google DeepMind’s AlphaEvolve shows how far this can reach. When it was unveiled in May 2025, Google reported AI-generated improvements that recovered an average of 0.7% of its worldwide compute resources, reduced Gemini’s training time by 1%, and contributed a verified circuit modification to an upcoming TPU. Google has since expanded its use well beyond those first results. AI is helping improve both the software and hardware used to build AI.

From better answers to lasting improvement

Stanford’s CS329A: Self-Improving AI Agents course, whose lectures are now available online, connects several mechanisms behind this shift: spending more compute exploring solutions, checking results with verifiers, learning from execution feedback, and training on successful attempts.

These represent different forms of improvement. An agent can try several approaches and produce a better answer without changing the underlying model. Persistent improvement requires retaining something useful, whether in memory, reusable code or updated model weights.

Consider a coding agent that writes a solution, runs it, encounters an error and revises its approach. Fixing that error improves the current result. Capturing the successful method and using it to improve future attempts creates a learning loop. The crucial test is whether those gains transfer to unfamiliar problems.

For startups, this raises a useful question: does your product get better with experience, or does it start from scratch with a more capable model each time? A system that turns verified customer outcomes into reusable improvements could build an advantage that grows with usage, provided it has permission to use that data.

How fast can this go?

The potential for compounding progress is clear. Cheaper inference enables more experiments. Better coding agents help researchers execute those experiments. Successful experiments can produce better models and further efficiency gains.

There is now evidence beyond code-generation statistics. In an independent assessment published on 5 November, METR estimated that Anthropic’s benchmark scores (measured on AECI, Anthropic’s version of the Epoch Capabilities Index) have grown at roughly 1.9 times their previous rate since October 2025, with a 95% confidence interval of 1.5 to 2.8. METR’s best guess is that around 80% of the current acceleration is due to AI-assisted R&D. It stressed substantial uncertainty, and noted that Anthropic redacted a majority of the evidence METR found most informative.

That does not establish an unlimited growth curve. Faster coding, faster research and faster capability improvement are different measures. If coding takes half a project’s time, doubling coding speed reduces total project time by only a quarter, assuming everything else stays unchanged. Testing, experiment runtime and deployment become the next constraints.

“More tokens equals better performance” also has an economic limit. Additional optimisation is worthwhile while the expected gains exceed the cost of generating, testing and maintaining changes.

Even with these constraints, founders should plan for a moving frontier. Products that are uneconomic today could become viable, while advantages built around current model limitations could disappear.

The risks begin before AI becomes autonomous

The immediate danger is optimising the wrong thing extremely well. An agent pursuing throughput might improve benchmark performance while introducing failures on unusual inputs. Passing a test suite establishes that the software passed those tests, not that it is universally correct.

This makes the evaluator as important as the agent. Objectives need constraints on security, quality and reliability, supported by independent tests that the optimiser cannot simply tailor itself to. Longer tasks also need end-to-end evaluation: an agent can perform individual steps well and still fail to complete the overall job reliably.

There is a risk of shared blind spots. If similar models write the code, generate the tests and review the results, agreement can create false confidence. Asking the same model to check its work in another conversation is a weak substitute for independent verification.

Human supervision can become less meaningful as the volume of changes grows. Someone may retain the authority to stop a system while losing the ability to understand when intervention is necessary. Effective oversight requires reproducible experiments, bounded permissions, inspectable evidence and tested rollback.

Competitive pressure compounds the problem. Companies benefit from accelerating research, even when slowing down to evaluate a new capability would be prudent. The concern is that development could outpace our ability to assess its consequences, and the warning lines are no longer hypothetical. Anthropic’s own Responsible Scaling Policy treats a doubling in the rate of progress, substantially attributable to AI-driven R&D, as a key risk threshold, and METR’s estimate puts that line within its confidence interval. Reproducible experiments, bounded permissions, inspectable evidence and tested rollback need to be standard practice, and independently checked, before fully autonomous research becomes commonplace, not after.

Where startups can build

As an early-stage investor, I see opportunities in three layers.

The first is optimisation itself. Together’s results suggest that databases, networks, cloud workloads and simulations can be continuously improved against measurable customer outcomes. Specialist engineering agents can do the same for bounded problems, where domain expertise and reliable evaluation make improvements demonstrable.

The second is the trust layer around that optimisation. Someone needs to prove that agent-generated changes preserve correctness, security and reliability. Someone also needs to provide the experiment infrastructure underneath: realistic workload testing, reproducible environments, cost controls and safe deployment. If the evaluator matters as much as the agent, this layer could prove as valuable as the agents themselves.

The third is new consumer experiences. Cheaper inference makes personalised learning, creative tools and AI-native games economically viable in ways they were not a year ago.

For Israeli founders, systems engineering and security are particularly interesting entry points. A small team can build valuable infrastructure around evaluating, controlling and improving AI systems without training a frontier model. Keewano, a Remagine portfolio company where I sit on the board, is one example: it is building a database designed for AI agents’ reasoning and analytics, infrastructure that becomes more relevant as more of the work is done by agents.

The defensibility question becomes sharper, however. Why will customers keep using your product as the underlying models improve? Representative workload data, trusted integrations and accumulated evidence of reliable performance are stronger answers than a clever prompt. The ability to learn from real outcomes could strengthen that advantage further.

At Remagine Ventures, I would ask founders three questions: what can your system improve, how do you prove it, and who decides when it can act?

AI helping build its successor could compress development cycles and unlock entirely new products. It could also compress the time available to discover mistakes. Building the mechanisms to verify and control that progress may be one of the biggest opportunities it creates.

• Author
• Recent Posts

Follow me

Eze Vidra

Co Founder and Managing Partner at Remagine Ventures

Eze Vidra is the founder of VC Cafe and the co-founder and managing partner of Remagine Ventures, a pre-seed fund investing in ambitious founders at the intersection of AI, technology, entertainment, gaming, and commerce with a spotlight on Israel.

He is a former General Partner at Google Ventures (GV) in Europe, former head of Google for Entrepreneurs in Europe, and founding head of Campus London, Google's first startup hub. Eze writes on Israeli tech, venture capital, artificial intelligence, and founder strategy.

He is also the founder of Techbikers, a nonprofit that brings together the startup ecosystem on cycling challenges in support of Room to Read.

Follow me

Latest posts by Eze Vidra (see all)

• When AI Builds AI - November 7, 2026
• Weekly Firgun Newsletter – November 6 2026 - November 6, 2026
• Israel’s AI Infrastructure stack - November 5, 2026