TL;DR. A self-driving company is one where people choose the destination and AI agents drive the steps: research, outbound, code, reviews, support triage. The term comes from Replit CEO Amjad Masad. The way there runs through agent loops: goals with a verifiable endpoint, context access and an escalation rule. Start in engineering, then let adoption pull its way through the company.
Self-driving company is the term of the moment, and for once there's a number behind the hype. Replit reports, cohort-controlled, 2.9x more code per engineer. The point isn't the increase, though. The point is what stayed flat: review latency, reversion rate, incidents. The usual trade-offs did not occur.
We build AI agents for product and GTM teams, inside Swiss software companies with 20 to 500 people. The Replit post describes the destination this work leads to. So let's take it apart: what the term means, what it doesn't mean, and how you get there without being an elite engineering company.
What you'll learn
- The definition of a self-driving company, in one quotable paragraph.
- The agent loop: the working pattern that makes self-driving possible.
- The sequence: why engineering comes first and how GTM becomes loop-ready after.
The thesis: self-driving does not mean empty of people
«A self-driving company is not one without people. People still choose the destination.» That's how Amjad Masad defines it. People decide which problems matter, make the hard trade-offs and take responsibility. Agents execute the steps. Mistake self-driving for peopleless, and you build the wrong thing.
🧨 What is a self-driving company?
A self-driving company is an organisation where AI agents independently take over work steps across every function: investigating incidents, reviewing pull requests, triaging support tickets, researching sales accounts, analysing business data. People set the goals, check the results and decide on questions of judgement. The term comes from the Replit blog post «The Self-Driving Company» from July 2026.
What matters about the Replit case: it's not a vision, it's an operations report. Escalated support tickets close 60 per cent faster. 30 per cent of human review time is saved because an agent assesses risk and only pulls in a second human when needed. And the strongest line of the post isn't about output: «People don't feel automated. They feel promoted.»
🛠️ How to get there: four steps
- Start in the verifiable domain. A bug is measurably wrong; a marketing draft is only debatably bad. That's why the rebuild starts in engineering, where tests and CI give every agent output a hard verdict. Our piece on vibe coding for founders shows what that feels like.
- Wire up the context. Without access to the systems that run the company (repo, CRM, wiki, ticketing, data warehouse), no agent can self-drive. Replit gave its agent access to GitHub, Linear, Notion, Slack and Zendesk, behind access policies and audit logging. The agent is only as good as its context.
- Define agent loops. An agent loop is a goal set by people, a verifiable endpoint, context access and an escalation rule for questions of judgement. No loop without a measurable endpoint: otherwise the agent produces activity, not results.
- Let adoption pull, don't push. At Replit the way of working spread because everyone watched engineers tag the agent in Slack. Watching convinces better than any training. That's why agents belong in the public channel, not just in the IDE.
🤖 Why context beats tools
Replit cancelled a seven-figure SaaS solution because its internal agent, running on its own context, was better, and beat specialised tools at a tenth of the cost. The lesson for the mid-market is not «build everything yourself». The lesson is: agents that see your context beat generic tools that don't. In the end you want to own agents instead of renting tools.
Does this work in GTM too?
Yes, once you make GTM measurable. The fair objection: code is verifiable, marketing gut feeling is not. The answer is measuring instead of guessing: buyer-journey tracking, buying signals and clean CRM data give GTM loops the same verifiable endpoint that tests give code. That is exactly what Autonomous GTM builds: a go-to-market that produces pipeline without extra headcount.
Replit itself provides the proof that agents don't stop at engineering: the sales team uses the agent for lead research with internal knowledge, account executives to prepare customer conversations, the support team for playbook answers with escalation summaries. Engineering was the beginning, not the end.
Which autonomy level is your company driving at?
The autonomy levels transfer the logic of self-driving cars to companies: from level 0, everything manual, to level 4, self-improving agents. They show, per function, how far agents carry today and where the next step lies. GTM and engineering rarely drive at the same level.
- Level 0, manual. Tools yes, but every work step runs through human hands.
- Level 1, assisted. Agents suggest, people execute. Autocomplete, text drafts, reply suggestions.
- Level 2, delegated. Agents complete clearly scoped single tasks. A human checks every result.
- Level 3, autonomous in the loop. Agents work in loops with a verifiable endpoint and an escalation rule. The human reviews at the endpoint, not every step. Replit's engineering works here.
- Level 4, self-improving. Agents measure feedback, propose improvements and validate them with benchmarks and A/B tests. Replit's continual learning system is the example.
The diagnostic question for your next leadership meeting: which level is your GTM driving at, which your engineering, and what's missing for the next one? Usually the answer is not a model, but context, a measurable endpoint or an escalation rule.
🎢 Outro
✅ What works. Agent loops with a verifiable endpoint: migrations, reviews, incident triage, lead research, support playbooks. Starting in engineering, spreading through Slack.
❌ What doesn't work. Self-driving without context access. Vague goals without measurable endpoints. And trying to rebuild Replit's infrastructure when software infrastructure isn't your core business.
⚠️ Warning. Self-driving doesn't mean no problems, it means new problems: more output needs new review patterns, system access needs policies and an audit trail. Skip the security foundation and you build an incident, not a system.
Back to the number from the start: 2.9x more code is not the result of a tool, but of a rebuild. People choose the destination, agents drive, and the quality metrics stay flat because the checking was built in too. That's the standard every agent initiative should be measured against.
Replit rebuilt itself. We build this into your company and hand it over: Autonomous GTM for the market side, Autonomous Coding Agents for the product side. Or tell us in 30 minutes, founder to founder, where you're stuck.
FAQ
What is a self-driving company?
A company where AI agents independently take over work steps across every function, while people set the goals, check the results and decide on questions of judgement. The term was coined by Replit CEO Amjad Masad in July 2026.
Does a self-driving company replace its people?
No. Replit's formula: «People don't feel automated. They feel promoted.» The work shifts from executing to directing: setting goals, checking results, holding direction. People who think in outcomes gain value.
Where is the best place to start?
In engineering. Code is the most verifiable domain: tests and CI give every agent output a hard verdict. From there the way of working pulls its way into support, sales and marketing, as long as the agents work visibly in Slack.
Does this work in GTM, where results are less measurable?
Yes, but only after the measurability work: buyer-journey tracking, buying signals, clean CRM data. They give GTM loops the verifiable endpoint that tests give code. Without a measurement foundation, GTM automation stays spray-and-pray.
What is an agent loop?
A goal set by people, a verifiable endpoint, access to the necessary context and an escalation rule for questions of judgement. Not to be confused with the gtm.science loop DECODE, SHAPE, AMPLIFY, EVOLVE: one is a working pattern for agents, the other a strategy framework.
