AI Concepts

Self-Driving Company

20 min read
On this page (21)

Definition: A self-driving company is an organization in which AI agents do much of the step-by-step work, while people choose the goals, make the trade-offs, and stay accountable for the outcome. Replit introduced the term in July 2026. In Replit's words, a self-driving company "is not one without people. People still choose the destination."

TL;DR: Replit coined the term on July 16, 2026. Agents take a goal, do the work, check it, and escalate when human judgment is needed. People set direction and approve. Replit predicts every Fortune 500 company will be self-driving by the end of 2030, a forecast and not a measured trend. Add an approval step to one workflow →

Think of a car with a capable driver-assist system on a long highway trip. You still pick the city, the route, and when to stop for fuel. The system keeps the lane and the distance, and it hands control back when the road gets hard. A self-driving company applies that split to work: people pick the destination, agents handle the driving, and the hand-back moments are designed in advance.

Where the Term "Self-Driving Company" Came From

Replit, the AI app-building company led by CEO Amjad Masad, named the idea in a post on July 16, 2026, by Amjad Masad and Scott Kennedy. The post describes how agents spread through Replit's own teams and ends with a plan to offer the same way of working to customers. Coverage and follow-up posts came over the next four months.

Date Event Who Source
Jul 16, 2026 "The Self-Driving Company" names the idea and reports Replit's internal results Amjad Masad and Scott Kennedy, Replit Replit blog
Aug 7, 2026 Interview: "how do you remove the bureaucracy?" and "a CEO is a kind of glorified router" Masad, interviewed by Casey Newton Platformer (news interview)
Oct 15, 2026 Talk at TechCrunch Disrupt in San Francisco: "the closest analogy is a 'self-driving company'" Masad TechCrunch, Oct 16, 2026 (news report)
Oct 26, 2026 "The Rise of the Self-Driving Company": the Coase framing, Replit for Business, and the Fortune 500 prediction Amjad Masad, Replit Replit blog
Oct 28, 2026 "Replit's Architecture Behind Secure Agent Work": the Agent Habitat and four security principles Luis Héctor Chávez, Replit Replit blog
Nov 7, 2026 "How To Build The Future" with Paul Graham: Replit as "customer zero" and as "the platform for the transformation" to self-driving companies Masad and Paul Graham, Y Combinator YouTube, at 28:56

The July post makes the core claim in plain terms. Agents at Replit "investigate production incidents, review pull requests, answer questions, analyze business data, triage support tickets, research sales accounts," and help improve Replit's own product. Replit calls the result "an expanding system of agents operating across the company: taking goals from people, gathering context, performing work, checking the results, and escalating when human judgment is needed."

The October post adds the economic frame. It opens with Ronald Coase's 1937 question, "Why do firms exist?", and his answer, transaction costs. Replit's argument is that companies carry their own coordination costs, "management, meetings, departments, reporting, and bureaucracy," and that agents can take on more of that coordination. Its definition: "people define the goals, exercise judgment and taste, and remain accountable. Agents use the company's knowledge and systems to execute toward those goals, with clear boundaries for when to act and when to seek approval."

In the Y Combinator interview, Masad described the idea as the future of how firms are organized, called Replit "customer zero" for it, and said he thinks Replit is "the platform for the transformation" to self-driving companies (quoted from the auto-generated transcript).

What Replit Reports From Its Own Teams

These are Replit's own figures from the July 2026 post. They are self-reported and not independently audited.

Area What Replit reports Period
Code output Lines of code contributed rose 5.8x, and 2.9x for a consistent cohort of authors Early January to late June 2026
Code review Review latency flat. The agent assesses risk and calls in a second human reviewer only when needed, saving "30% (and growing)" of human review time Same period
Quality Pull request reversion rates and incidents opened stayed flat Same period
Support "A self-driving support team closes the hardest tickets (those escalated to humans) 60% faster" Not stated
Software buying Replit "churned a seven-figure SaaS solution" for an internal app built on Replit Not stated

How a Self-Driving Company Works

Every description from Replit has the same shape: a person sets the goal, an agent does the legwork, and a person decides at the points that carry risk. The loop below shows one unit of work, from the first signal to the record that the next run can learn from.

  1. Signal. Work arrives: a support ticket, a monitoring alert, a sales lead, a question about the data.
  2. Agent triage. An agent pulls the context a person would have gathered by hand. At Replit, per the TechCrunch report, a new support ticket is first examined by an agent that looks at GitHub issues, the codebase, and Linear tickets.
  3. Proposed action. The agent drafts the reply, the fix, or the report, and states what it wants to do next.
  4. Human approval. A person reviews the proposal where the cost of a mistake is real. Masad at Disrupt: "there's human judgment in the middle, and we don't think that's going away anytime soon." A rejected proposal goes back to the agent with the reason, and the revised proposal returns for approval.
  5. Execution. The agent or an automation carries out the approved step.
  6. Record. The run leaves a trail. Replit's October post asks that a correction to an agent "be reviewed and made available to improve future work."

The same loop seen as a conversation, using the support example from Masad's Disrupt talk:

The Habitat Behind It

Replit's October posts argue that agents need a controlled place to work, which Replit calls an Agent Habitat: "an environment where they can use the company's knowledge and systems, coordinate with people and other agents, and act within the boundaries the company controls." The habitat is Replit's product term, but the four principles generalize to any team that delegates work to agents.

Replit's principle (Oct 28, 2026) What it means in practice
"Give agents only the authority they need" Control what agents can access, which actions require approval, and how permissions carry through delegation
"Contain failures and respond to vulnerabilities" Isolate execution, protect credentials, and patch where the work runs
"Expand access while keeping production changes accountable" Let more people contribute, with production decisions still subject to review
"Keep delegated work accountable and data governable" Make agent activity visible, support investigations, and set data boundaries

Levels of a Self-Driving Company: The SAE Driving Analogy

The name borrows from cars, so the car standard is a useful ruler. SAE International's J3016 recommended practice, first published in 2014 and revised in 2021, defines six levels "from Level 0 (no driving automation) to Level 5 (full driving automation)" (SAE, May 3, 2021). The official SAE chart splits them into two groups. At Levels 0 to 2, "you are driving whenever these driver support features are engaged." At Levels 3 to 5, "you are not driving when these automated driving features are engaged."

The mapping from car levels to company examples below is an analogy made by this page. Replit does not publish a level scale, and SAE J3016 covers vehicles only.

SAE level (J3016 name) Who drives, per the SAE chart Company analogy (this page's mapping) Example
0: No Driving Automation You drive. Features give warnings and momentary help Software stores data and sends alerts. People do every step A shared inbox where a person reads and answers each ticket
1: Driver Assistance You drive. The feature helps with steering or speed AI helps with one step inside one tool AI drafts a reply that a person edits and sends
2: Partial Driving Automation You drive and must constantly supervise Agents chain several steps, and a person watches every run An agent triages, researches, and drafts, and a person checks each draft
3: Conditional Driving Automation You are not driving, but "when the feature requests, you must drive" Agents run a workflow end to end and hand back at set points Agents work tickets and escalate "when human judgment is needed," Replit's July description
4: High Driving Automation The feature drives under limited conditions and will not require you to take over Agents own a fenced, reversible domain. People set goals and audit Routine tagging and routing of incoming requests, with spot checks
5: Full Driving Automation The feature drives everywhere, in all conditions A company with no people in the loop None. Replit's definition rules it out

On this reading, Replit's definition sits around Levels 3 and 4: agents act alone inside boundaries, and the company decides in advance when they must hand back. A company does not have one level. Each workflow has its own, and the same team can run one process at Level 2 and another at Level 3.

The car world also offers a caution about the word itself. The U.S. National Highway Traffic Safety Administration does not use "self-driving" for higher levels of automation, because the term "describes a vehicle's state of operation but not necessarily its capabilities" (NHTSA). The same care applies here: "self-driving company" describes how work flows, not how much a company can safely leave to agents. For agent-level scales, see Agent Autonomy Levels.

What a Self-Driving Company Is Not

  • Not a company with no people. Replit's first definition says so directly, and its October post says people "remain accountable." Replit describes the change for employees as a promotion, in its words: "Self-driving turns doers into directors."
  • Not full autonomy. Agents act inside boundaries the company sets, including "when to act and when to seek approval." Level 5 on the analogy above is outside the definition.
  • Not one product. Replit sells a product for it, Replit for Business, but the idea is an operating model. Any team can apply the loop to one workflow with the tools it already has.
  • Not a measured industry state. The Fortune 500 prediction is Replit's forecast, and the productivity figures are Replit's own numbers.
Term What it describes Scope Where people sit Read more
Self-driving company An operating model where agents do the steps and people choose goals and approve The whole organization Set direction, approve, stay accountable This page
Autonomous agents An AI system that plans and acts toward a goal with little supervision One agent Set the goal and limits Autonomous Agents
Agentic workflow One process in which an agent decides and automation acts One workflow Review the cases the agent cannot settle Agentic Workflows
Human-in-the-loop A design pattern: an agent proposes, a person approves, then the action runs One step At the approval gate Human-in-the-Loop
One-person company A business run by one founder, with AI and software doing work a team once did A small business One person does all the deciding One-person companies
AI-native company A general label for a company built around AI from the start, in product and operations. No single source defines it Company design Varies No dedicated page

The difference that matters most is scope. An autonomous agent is a component, an agentic workflow is a process, and a self-driving company is what you get when many such processes run across functions with clear hand-back rules. A one-person company and a self-driving company can overlap: a solo founder can run several workflows this way, but the term from Replit describes firms of any size.

Risks and Open Questions

The same sources that promote the idea also name its limits. The table keeps to what they state.

Risk What can go wrong What the sources say Common safeguard
Accountability No one can say who answers for an agent's action Replit: people "remain accountable," and agent activity must be visible to "support investigations" A named owner per workflow and a log of every decision
Error cost An agent changes live data or systems In July 2025 a Replit agent deleted data from a user's production database, which Masad called "Unacceptable and should never be possible" (Masad on X, reported by The Register). Replit then launched separate development and production databases (Replit, Jul 21, 2025) Separate test and live data, approval before irreversible steps
Review load Agents produce more work than people can check Replit kept review latency flat by letting an agent assess risk and call in "a second human reviewer when necessary" Risk-based review: agents screen, people review the high-risk cases
Security Broad connections and prompt injection Replit: "If a task only requires reading but the connection also permits writing, a mistake or prompt injection can have consequences beyond the task itself" Least privilege, credentials kept out of the agent's reach, isolation
Spillover A mistake reaches past the company Replit's October post points to OpenAI's report on the Hugging Face incident (Aug 26, 2026) Approval rules for anything that leaves the company
Evidence Results do not transfer Replit's figures are self-reported from one company that builds agent tools Measure your own baseline before and after

Open questions remain. Replit's own framing names "a false choice" between keeping AI "in a cage" and letting it "loose in the wild," and argues that controls solve it. Whether those controls hold at a large company, how review time scales as output grows, and whether the 2030 prediction comes true are all things the coming years will show. Treat any single forecast with caution, and see AI Guardrails and Prompt Injection for the controls in more depth.

How Self-Driving Companies Connect to Taskade

Taskade does not run a whole company for you, and no product removes the need for people to own the decisions. What a team can do today is run one workflow in the self-driving pattern. A Taskade Genesis app holds the workflow's data in workspace projects, a Taskade AI agent does the triage and the drafting, and an automation runs the steps.

The Wait for approval step is the hand-back point. It pauses an automation run until a workspace member approves or rejects. Approve continues the run, and Reject stops it. If no one answers within the wait you pick (1, 3, or 7 days), the Then setting decides. Later steps can read the decision, who made it, and when, so you can write each decision to a log project. That keeps a person in charge of the trade-off, the core of the self-driving pattern.

Plan facts, checked 2026-11-12: Free includes 3 Taskade Genesis apps, 1 agent, and 10 automation runs per billing period, and starts with 0 AI credits. On Free, you earn free AI credits with referrals and invites, and you can upgrade any time. Premium connectors, including the Gmail, Slack, and Microsoft Teams approval actions, need a paid plan. Pro is $10/mo billed annually.

What You Would Build in Taskade

Picture a support triage desk for a small software team. Each new ticket lands in a project. An agent reads it, searches your help articles and past tickets, and drafts a reply with a likely cause. An automation then sends the draft to the on-duty teammate through a Wait for approval step. Only after the teammate approves does the next step send the reply, and a final step logs the decision.

In Taskade you describe that in one prompt, and Taskade Genesis builds the app with its ticket data, the triage agent, and the approval automation. The agent does the legwork, and your team keeps the decision. Describe yours and start building →

Frequently Asked Questions About Self-Driving Companies

What is a self-driving company?

A self-driving company is an organization in which AI agents do much of the step-by-step work while people choose the goals, make the trade-offs, and stay accountable. Replit introduced the term on July 16, 2026, and wrote that it "is not one without people. People still choose the destination."

Who coined the term "self-driving company"?

Replit did, in "The Self-Driving Company," a July 16, 2026 post by CEO Amjad Masad and Scott Kennedy. Masad expanded on it in an August 2026 Platformer interview, at TechCrunch Disrupt in October 2026, in the October 26 post "The Rise of the Self-Driving Company," and in a November 2026 Y Combinator interview with Paul Graham.

Does a self-driving company mean no employees?

No. Replit's definition says people define the goals, exercise judgment and taste, and remain accountable. Agents carry the work toward those goals within boundaries for when to act and when to seek approval. Replit describes the change for its staff as a promotion from doing the steps to directing them.

Is a self-driving company the same as an autonomous company?

Not in Replit's sense. "Autonomous company" has no standard definition and often suggests a firm that runs with no people. A self-driving company keeps people at the decision points. On the SAE driving analogy, it sits around Levels 3 and 4, where agents hand control back at set points, not at Level 5.

What is Replit for Business?

Replit for Business is the offering Replit presents in its October 26, 2026 post as its foundation to help other companies become self-driving. Replit's October 28 post describes its Agent Habitat, a controlled environment where agents use company knowledge and systems within boundaries the company sets, with access controls, approval rules, isolation, and activity logs.

Will every Fortune 500 company be a self-driving company by 2030?

That is Replit's prediction. Its October 26, 2026 post states: "By the end of 2030, every company in the Fortune 500 will be a Self-Driving Company." It is a forecast by a company that sells tools for the shift, not a measured trend.

How can a small team start?

Pick one repetitive workflow, such as support triage or invoice review. Let an agent gather context and draft, add an approval step before anything that sends a message, changes live data, or moves money, and log each decision. In Taskade, the Wait for approval step pauses an automation until a teammate approves or rejects.

What are the main risks of a self-driving company?

Unclear accountability, costly errors on live data, more output than people can review, and security exposure from broad connections and prompt injection. The common safeguards are a named owner per workflow, least-privilege access, separate test and live data, risk-based review, and an approval step before irreversible actions.

Further Reading