Cognition AI's Devin coding agent helped factor RSA-260, a 260-digit number, in what researchers estimate cost around $400,000 in computing resources. The achievement made headlines as a cryptographic breakthrough, but the deeper insight reveals how AI coding agents are reshaping the economics of technical work. Instead of replacing engineers, Devin amplified what one skilled person could oversee and direct.

What Actually Happened With RSA-260?

On August 13, 2026, Cognition researcher Eric Lu pointed Devin at a practical engineering challenge: build a fast graphics processing unit (GPU) replacement for a CPU-based component in CADO-NFS, a well-known factoring toolkit. Two hours later, he asked it to handle parameters for RSA-250. Then he went to bed. Seven hours after that, Devin had delivered a working version.

The factorization itself used established mathematics, not new algorithms. Lu's technical account emphasizes "essentially no algorithmic advancements," only "good old performance engineering." What changed was the cost and the human effort required. The previous record, RSA-250, consumed roughly 2,700 CPU core-years in 2020. The RSA-260 run used about 13.5 GPU-years of computing power, which Lu estimates cost approximately $400,000 at market rates. Cognition says its GPU lattice siever reduced costs by roughly tenfold compared to the previous public state of the art.

The timing created a neat narrative: Cognition announced a $2 billion funding round at a $48 billion valuation just days after the factorization was revealed. But Lu had started the work weeks before the funding report, and the RSA-260 run used spare cluster capacity that Cognition says had no marginal cost. There is no public evidence the factorization was staged around the financing.

How Did One Engineer Supervise an AI Agent on Such Complex Work?

Lu's role throughout the project reveals how AI coding agents are changing the division of labor in technical work. He did not disappear from the process; instead, his job moved upward in scope. Rather than writing every kernel and operating every job himself, Lu spent more time deciding what the swarm of AI agents should do, whether their results could be trusted, and when to change course.

Across 233 Devin sessions, Lu sent 82,702 words in 3,328 messages. The agents also started 101 child sessions themselves, with 36 receiving no direct intervention from him. This is neither fully autonomous work nor simple autocomplete. Lu set priorities and scope, stopped useless measurements, pointed out directions the agents had missed, and imposed comparable experimental frameworks. The agents handled measurements, cluster operations, debugging, and large amounts of optimization work.

Lu estimates the project took about three weeks from first prompt to factors. He notes this replaced what would probably have been months of work by a team of specialists, though Cognition's comparison has not been independently audited. The division of labor is revealing: one strong engineer now supervises an amount of implementation that used to justify several salaries.

Steps to Understanding How AI Agents Amplify Engineering Productivity

• Structured Problem Decomposition: Lu emphasizes that CADO-NFS mattered because it already broke the factoring problem into stages with stable interfaces and reference implementations. The agents became more confused as their code moved further away from upstream CADO-NFS, suggesting that human-designed structure is essential for AI agents to work effectively.
• Continuous Human Oversight: Lu did not hand off the project and wait for results. He maintained a technical lead role, catching the agents when they abandoned promising work too early, ensuring measurements and benchmarks remained coherent, and redirecting effort when needed.
• Leveraging Existing Infrastructure: The project succeeded partly because it built on decades of mathematics, GPU hardware, Cognition's cluster infrastructure, and the underlying models powering Devin. The agents worked best within established systems rather than inventing new ones.

What Does This Mean for the Future of Software Engineering?

The RSA-260 factorization hints at a larger commercial shift. If one strong engineer can supervise an amount of implementation that previously justified several salaries, the economics of software engineering, technical consulting, and parts of scientific research could change significantly. This is not the same as saying AI will replace engineers; rather, it suggests that skilled engineers will become force multipliers, directing AI agents rather than writing every line of code themselves.

Cognition's business model reflects this bet. The company post-trains models rather than merely exposing another company's API. SWE-1.7 was trained from a Kimi K2.7 base, and Cognition's newer SWE-2 was post-trained from Kimi K3. Value is built in post-training, orchestration, tools, memory, permissions, evaluation, and the customer relationship, not just in the underlying base model.

The authorship question is thorny. The factorization rests on decades of mathematics, CADO-NFS, GPU hardware, Cognition's infrastructure, the models operating underneath Devin, and Lu's judgment. Lu himself says he cannot cleanly apportion credit among the participants. Yet marketing has fewer philosophical difficulties; the achievement gets attached to the product.

What remains clear is that the cost of expertise has shifted. Old mathematics made cheaper can become economically new. Nobody needs a new theorem if a known method suddenly costs one-tenth as much to deploy. For organizations building software, that shift may matter more than the cryptographic headline.