The Self-Driving Company

One-line job: Understand what Replit actually shipped to make agents do routine work company-wide — and what it would take to copy the pattern honestly. Audience: Operators and managers wondering whether “agents run the company” is vision-deck fluff or an operating model. Not for: Anyone looking for a promise that agents eliminate headcount — the source essay explicitly rejects that reading. Last verified: 2026-08-12 Evidence weight: documentation-verified

“The self-driving company” is Replit CEO Amjad Masad’s term — coined in a July 16 essay — for an organization where people set the destination and AI agents do the driving. One correction up front, because the phrase invites misreading: Masad does not mean a company without people. People choose which problems matter and own outcomes. What changes is who performs each step.

The anchor line from the essay:

“People don’t feel like they’ve been automated. They feel like they’ve been promoted.” — Amjad Masad, Replit, 2026

The field report, not the vision deck

What makes this worth attention is that Replit published numbers rather than adjectives — with the caveat that they are self-reported by a vendor selling an agent platform:

• 2.9× code output per engineer, holding cohort constant (5.8× total lines January–June including hiring effect)
• ~30% of human PR-review time saved
• PR reversion rates and incidents flat — speed did not cost quality
• 60% faster closure on the hardest human-escalated support tickets
• Internal agent matched market-leading vertical tools at roughly a tenth of the cost — and the company churned a seven-figure SaaS contract after an internal build won

Treat these as one AI-native company’s field report. No independent audit exists. But flat incident rates alongside 2.9× output is the number that should make skeptics look twice, because it addresses the standard objection (agents create cleanup debt) with data.

How it actually works: loops plus governance

The mechanism is loops: goals set by people, agents given governed access to needed systems, verifiable criteria to check progress, and escalation to humans when judgment is required. Every employee gets a manager agent that spawns fleets working on verifiable tasks — CSS migrations, localization, flaky-test maintenance. One networking bug reportedly got cracked by a swarm.

The unglamorous part is the prerequisite. Replit locked its internal agent behind access policies, token proxies, audit logging, and a ZeroTrust network before giving it access to GitHub, GCP, Azure, Linear, Notion, Slack, and Zendesk. They also built what the essay calls a semantic layer — canonical metrics and sources of truth — described as “the first act of governance for an AI-native company.”

Adoption then spread out of engineering through Slack, pull-not-push: non-engineers watched engineers tag the agent with tasks and tried it themselves.

The honest gap

Here’s the part most coverage skips: Replit is AI-native. No adoption debt, an in-house agent platform, young integration-friendly systems. For a typical enterprise carrying a twenty-year-old ERP and a regulated data estate, Snowman Labs’ roadmap is more realistic — staged: assisted individuals first, engineering as the proving ground (its work is the most verifiable), then cross-functional extension.

Masad himself, in a Platformer interview, frames the CEO role as “a kind of glorified router” whose routing functions should mostly be automated — and predicts companies get smaller even as net jobs grow. That’s a prediction, not data. But it tells you which direction the person running this experiment thinks it goes.

The call

1. Check your prerequisites before your ambitions: is your work verifiable? Can agents be given governed access to where work happens?
2. Build the governance layer first — access policies, audit logging, ZeroTrust — then grant access, never the reverse.
3. Start with engineering as the proving ground; it has the clearest success criteria.
4. Discount Replit’s numbers appropriately — self-reported, AI-native, single company — but don’t dismiss the flat incident rates.
5. If your org lacks verifiable success criteria for its work, fix that before buying any agent platform. The loop needs something to check.