Marc Andreessen, the venture capitalist who cofounded Andreessen Horowitz, used a post on X to confirm his firm is backing Cognition, the maker of the coding agent Devin, for a second time. The disclosure doubles as an argument: Andreessen says the software industry he described in 2011 is now compounding faster than he predicted, and his money is the proof he wants readers to notice.
That framing matters because Andreessen is not a neutral observer here. He is a returning investor in the company whose numbers he is citing, which means every figure in the post also functions as marketing for a portfolio bet.
The core data point is internal to Cognition. Andreessen says Devin’s share of the company’s own production code rose from 13 percent a year ago to more than 90 percent now, a shift he credits to agents doing most of the actual coding work rather than assisting a human who does it. He extends that claim to customers: Mercedes-Benz reportedly turned an eight-month COBOL migration into eight days, Rivian’s test-generation velocity increased tenfold, and the Brazilian bank Itau now has Devin automatically remediate 70 percent of the security vulnerabilities it flags.
None of those figures carry independent verification. They are numbers Cognition and its investor chose to publish, not numbers audited by a customer or a third party, and the post does not address what “automatically remediated” means in practice for a bank’s compliance obligations.
Andreessen’s larger argument leans on precedent: he lists compilers, open source, and cloud computing as prior leaps that were each supposed to shrink the software engineering profession and instead expanded demand for it. He attributes the pattern to Milton Friedman’s view that human wants are effectively unlimited, so cheaper software production creates more work rather than less. It is a reasonable historical pattern, though citing three past technology shifts that didn’t eliminate jobs is not the same as demonstrating that agentic coding will follow the identical curve.
The post also revives Andreessen’s own 2011 essay, “Software Is Eating the World,” noting that technology’s share of market capitalization among the world’s ten largest companies rose from 31.5 percent to 94.4 percent over the intervening fifteen years. Reusing that framework here is a deliberate move: Andreessen is positioning Devin’s growth inside a thesis he has already been proven right about once, which makes a second bet easier to sell to readers than a cold pitch for an unfamiliar startup would be.
He also quotes Cognition’s chief executive Scott Wu, who argues engineers will “go from bricklayers to architects, focusing on the creativity of designing systems rather than the manual labor of putting them together.” That is a claim about what engineers will do next, not evidence about adoption speed or reliability, and it comes from the founder Andreessen is funding rather than from an independent source.
For operators evaluating coding agents, the practical test is not Cognition’s own production-code percentage but whether a specific team’s error rates and review overhead improve when an agent handles the bulk of commits. Teams piloting Devin or comparable agents this quarter should ask vendors for customer-side audit data, not headline percentages, before recalculating engineering headcount plans for 2027.
Marc Andreessen, cofounder of Andreessen Horowitz, posted these claims on X on September 9, 2026.
