How a Self‑Driving Company Tripled Code Output Without Raising Incidents

In a "self‑driving company" humans define objectives, priorities and bear final responsibility, while a coordinated network of AI agents performs information collection, task execution, result validation and routine collaboration. The model does not replace employees; it frees them from repetitive work to focus on strategic judgment, creative thinking and key decisions.

Productivity Gains at Replit

Over the past six months Replit’s engineers increased code output by almost three times. Code‑review latency remained stable, rollback frequency and production incidents did not rise, quality metrics improved, and release cadence accelerated. The most visible effect is the surge in code, but deeper changes occurred beneath the surface.

AI Agents in Action

Agents now investigate online incidents, review pull requests, answer questions, analyse business data, triage support tickets and research sales customers, continuously improving the Replit Agent system itself. This functions like a "master intelligence" that receives goals from people, gathers context, executes tasks, checks results and escalates to humans when judgment is required.

Infrastructure and Security

At the end of January the team rolled out internal infrastructure to experiment with agent applications, using an existing agent runtime framework, micro‑VMs and a remote file system so every engineer could orchestrate multiple agents in parallel. Access policies, token proxies, audit logs and a Zero‑Trust network secured the system, allowing agents to interact with GitHub, GCP, Azure, Linear, Notion, Slack, Zendesk and other services.

Measured Impact

From early January to late June the number of code lines submitted grew 5.8×; after removing the hiring effect, a fixed group of contributors still produced 2.9× more code. With team size doubling, per‑engineer output rose threefold.

Code‑review delays stayed flat because agents now participate in reviews, assessing risk levels and involving a second human reviewer only when necessary. This has saved roughly 30% of manual PR‑review time, a figure that continues to improve.

Quality did not suffer: PR rollback rates and new incident counts remained steady, while agents helped capture more defects and accelerated incident root‑cause analysis, shortening mean time to mitigate (MTTM).

Business‑Level Outcomes

Feature delivery to users increased, as tracked via Linear, leading to a noticeable rise in project completion rates. Agents also empowered sales, product, support, marketing and other functions: sales teams qualified leads with richer context, product managers performed release analyses, support agents resolved tickets 60% faster, and marketing drafted specifications from a single prompt.

Case Studies

Engineers used agents to complete a long‑stalled CSS migration, automate localisation, maintain flaky tests, and resolve a complex network issue involving PSC and file‑descriptor closures. A continuous‑learning system now analyses user feedback, proposes improvements, and validates changes through benchmarks and A/B tests, enabling the Replit Agent to self‑improve.

Build vs. Buy Shift

Internal agents have begun outperforming many industry‑leading products. Replit retired a seven‑figure SaaS solution in favour of a fully internal implementation that delivers better performance at a fraction of the cost. Similar gains were seen against vertical tools for alert triage and automated penetration testing.

Scaling Beyond Engineering

The model spread quickly to other departments, driven largely by Slack integration. Employees use agents to ask knowledge‑base questions, edit documentation, and generate presentations without waiting for engineering responses. Data teams provided a semantic layer over the data warehouse, allowing anyone to ask BI questions and receive reliable answers, freeing the team to tackle harder problems.

Product managers now run complex release analyses, sales developers enrich qualified leads, and support agents investigate issues and hand off escalations with contextual replies. Across the company, humans are not replaced by automation; instead they are promoted to commanders, focusing on outcome‑driven work.

Future Outlook

The self‑driving model is still evolving, with models continuously improving. Replit sees this as only the beginning of a broader transformation where AI agents become integral to every workflow.