A question has hung over the AI conversation for years: how far can agents actually change how a company operates? Do they stay a “coding assistant,” or do they rebuild the organization itself?
Replit CEO Amjad Masad gave a bold answer. His “Self-Driving Company” concept drew over 1.1 million views within a day of posting — and it is not talk. Replit’s last six months of internal practice produced genuinely hard numbers, and for the first time outsiders can see what happens to organizational efficiency when agents are woven into the company’s fabric.
The Data Doesn’t Lie: 3x the Code, Fewer Incidents
The most visible change hit engineering. From January to June 2026, Replit engineers’ committed lines of code grew 5.8×. Stripping out hiring effects, per-person output on the fixed team reached 2.9×.
The surprise is what happened to quality: code-review latency stayed flat, PR rollback rates and new incident counts did not rise — they fell. Team headcount doubled while per-person output tripled and quality improved. In conventional engineering management that combination is a near-impossible triangle.
The key: agents do not just write code — they joined code review, incident investigation, and root-cause analysis. Human reviews now run with an agent co-reviewer catching more defects; incident investigations use agents to localize root causes, and mean time to mitigation is shrinking. Human reviewer time is down 30% and climbing.
How It Works: From AI Assistant to Organizational Infrastructure
Step one: infrastructure first
In late January, Replit switched on internal agent-experiment infrastructure: an agent runtime framework, micro-VMs, remote file systems, plus access policies, token proxies, audit logging, and zero-trust network protection. Then the agents were connected to GitHub, GCP, Azure, Linear, Notion, Slack, and Zendesk. The moment agents gained cross-system context, productivity leapt — the single-tool agent and the cross-system agent are different species.
Step two: the agent-manager model
Every employee has a managerial agent that can spawn multiple sub-agents and orchestrate their loops on your behalf. Engineers design “loops” — dispatch a group of agents at a verifiable task and concentrate yourself on higher-level judgment. Replit’s CTO even used a swarm of agents to break a network problem that had dogged the team for months.
Step three: the self-improving loop
Replit’s AI team built a continuous-learning system: analyze user feedback, propose improvements, validate them with benchmarks and A/B tests. Replit Agent is improving itself — the most compelling “self-driving” case of all.
From Engineering to the Whole Company: a Natural Diffusion
Agent usage spread from engineering to the rest of the company mostly through the Slack interface — other teams watched engineers assign tasks to agents in Slack and started experimenting on their own.
Data: the team gave agents a semantic layer over the data warehouse, so anyone can ask business-intelligence questions and get reliable answers while the data team concentrates on the genuinely hard problems. Sales: agents find product-qualified leads, and because they can read internal knowledge — usage levels, project activity, contract details — outreach is sharply targeted; account managers generate customized communications and demo decks per customer. Marketing: one prompt turns engineering and product conversations and documents into draft product specs, no meeting attendance required. Support: resolution of the hardest tickets accelerated 60% — agents investigate, follow standard procedures, choose standard replies, or escalate to engineering.
In every case the human was not automated away — the human moved from executor to commander. That is the core logic of the self-driving company.
An Unexpected Side Effect: Rethinking Build Versus Buy
Once the internal agent system matured, Replit noticed something unexpected: its need for external tools collapsed. A seven-figure annual SaaS contract was cancelled because an internally built application performed better. A competing alert-triage tool of similar quality cost ten times their internal solution to run. An automated penetration-testing tool found fewer vulnerabilities at ten times the cost. The self-driving company does not just change internal efficiency — it rewrites technology procurement: many purchased tools can now be built in-house by agents, cheaper, more customized, and better.
The Frame: Three Layers of a Self-Driving Company
-
Cross-system agent access — agents reach GitHub, Slack, the data warehouse, and ticketing so they hold complete context.
-
Agent managers, not point tools — every employee commands a managerial agent that dispatches sub-agents for complex tasks; the human owns judgment and decisions.
-
Agents inside the verification loop — not just execution but checking, testing, and review, forming an execute → verify → optimize cycle.
Notably, Boris Cherny — creator of Claude Code — published an “AI adoption steps” table in the same window placing “self-driving company” at stage 3 (supervised autonomy) with touches of stage 4 (AI-native) in closed-loop scenarios. That maps precisely onto what Replit is showing.
This Is Not Only Replit’s Story
Replit, as an AI application-development company, enjoys natural advantages. But the self-driving-company frame is general: humans set direction, agents execute and verify, and the relationship between them is command — not replacement. Any organization with an internal knowledge base, cross-system collaboration needs, and repetitive workflows can start from engineering and diffuse into data, sales, marketing, and support. The catch: this is not solved by buying one AI tool. It requires design at the level of organizational architecture and infrastructure.
As Masad said: “This is just the beginning.” As model capability keeps climbing, the boundary of the self-driving company keeps moving outward. The next question may be the more unsettling one: when agents become a company’s default infrastructure, what happens to the manager’s role?
