Replit, the browser-based coding platform, says a fleet of internal AI agents helped engineers increase code output nearly threefold over six months without adding reverted pull requests or production incidents. The company laid out the numbers in a blog post called “The Self-Driving Company,” co-authored by chief executive Amjad Masad and engineer Scott Kennedy. Every figure in the post comes from Replit’s own dashboards. No outside party has reviewed the underlying data.
The core claim involves scale. Code contributed by engineers rose 5.8 times between early January and late June, Replit says. Part of that jump reflects hiring, so the company isolated a cohort of engineers who stayed on staff the entire period. Within that group, output rose 2.9 times per engineer, a combination Replit calls unusual for an engineering organization scaling headcount at the same time.
Replit built the system by connecting agents to Slack, Notion, Linear, GitHub, GCP, Azure, and Zendesk, gated behind access controls, token proxies, audit logging, and a zero-trust network. Agents now investigate outages, review pull requests before a human sees them, query the company’s data warehouse for business questions, and close support tickets against a standard playbook. Agent-assisted review has saved 30 percent of human reviewer time, Replit says, and its support team now closes escalated tickets 60 percent faster.
None of this is independently verified. The productivity figures come from the same internal systems whose adoption Replit is measuring, and the post does not define how the engineer “cohort” was selected, what counts as a code contribution, or whether the comparison window overlapped with the sprint crunch that precedes a major release. Replit’s core business is selling an agentic coding product, Replit Agent, so a case study showing its own staff thriving under agent supervision doubles as a sales pitch to the enterprise buyers evaluating that same tool.
The sharper data points involve cost rather than volume. Replit says it dropped a seven-figure SaaS contract after building an internal tool that outperformed it, and that a purchased alert-triage product matched its internal agent’s quality at ten times the price. A commercial penetration-testing tool, the company says, found fewer vulnerabilities than its homegrown version while charging ten times as much. If those comparisons hold up, they describe engineering organizations with enough internal capacity choosing to build agent infrastructure instead of buying narrow AI products.
That is the more durable story, independent of whether “tripled” survives outside scrutiny. Replit is positioning agents as company-wide infrastructure, threaded through sales, marketing, and support rather than confined to a coding editor. Engineering leaders evaluating vendor AI tools should not adopt Replit’s figures at face value. Instead, pick one recurring task, run an internal agent against the purchased alternative for thirty days, and compare cost and defect rates directly rather than trusting either side’s self-reported numbers.
Replit detailed these figures in a post on its official company blog, “The Self-Driving Company,” which does not carry a published date.
