What Is a Self-Driving Company? The Levels, the Missing Layer and How to Test One (2026)
Explainer · October 2026 · 8 min read
By Ashwin Gopinath, CEO
TL;DR
- A self-driving company is one where AI agents carry out most of the execution, such as investigating incidents, reviewing code, triaging tickets and analyzing data, while people choose the destination, make the tradeoffs and own the outcome. Replit CEO Amjad Masad named the idea in July 2026.
- The essays describe the driver. They say much less about the map. Agents that act across a company need the same current picture of it, permissions that follow every source, and a record of what they did.
- A self-driving car is tested in simulation long before it meets a real road. A self-driving company can be tested the same way: the Glass Company is a simulated 1,200-person company built and scored by Era by Eon, where Sentra found 495 of the 497 problems hidden in it.
- Sentra is the company memory and judgment under the agents. It reads email, chat, meetings, documents and systems of record, works out what is true now and who owns it, and raises what is going wrong to the right person, who approves before anything is sent.
See what your leadership team would know before Monday's review.
Watch it find 495 hidden problemsBook a demoWhat is a self-driving company?
A self-driving company is an organization where AI agents do most of the step-by-step work and people direct it. Agents investigate, draft, review, triage and analyze. People decide which problems matter, choose between options, apply taste and remain accountable.
The term comes from Amjad Masad's essay "The Self-Driving Company", published on 16 July 2026. He described how Replit's agents now investigate production incidents, review pull requests, answer product questions, analyze business data and triage support tickets. Replit reported that the same engineers nearly tripled their code output in six months, that agent-assisted review saves about 30% of human review time, and that support closes escalated tickets 60% faster.
Masad was explicit that the self-driving company is not a company without people. Speaking at TechCrunch Disrupt in October 2026, he put it plainly: "You need human judgment; you need taste." People move from doing every step to directing the work.
Where the idea came from, and why it is everywhere now
| Date | What happened |
|---|---|
| 16 July 2026 | Masad publishes "The Self-Driving Company", which passes a million views |
| 19 July 2026 | The AI Daily Brief devotes an episode to it |
| 16 October 2026 | Masad repeats the framing on stage at TechCrunch Disrupt |
| 26 October 2026 | Marc Andreessen writes it up for a16z in Let Chaos Reign, and Then Rein In Chaos |
Andreessen's piece adds the uncomfortable half. Agents spread through companies the way spreadsheets and email once did, from the bottom up, and the early state is chaos. He quotes Masad on agents running with "whatever credentials and data someone handed them", and on the result: "Nobody can see what they did." The a16z answer is a contained habitat with access rules and auditing.
That is the right diagnosis. It also points at the missing layer.
The levels of a self-driving company
Self-driving cars are described in levels, from driver assistance to full autonomy. Companies are moving through a similar ladder. This is a working framework, not an industry standard.
| Level | What agents do | What people do | Typical sign |
|---|---|---|---|
| 0. Manual | Nothing | Everything | AI is banned or unused |
| 1. Assisted | Answer when asked, one person at a time | Prompt, check, copy and paste | Everyone has ChatGPT, Claude or Copilot |
| 2. Partial | Run whole tasks inside one team's tools | Review each result | Agents triage tickets or review code in one function |
| 3. Conditional | Work across teams and systems, raise what needs a decision | Approve, redirect, own outcomes | One shared picture of the company that every agent reads |
| 4. Self-driving | Notice problems nobody asked about and start the fix | Set direction and approve what goes out | The company finds its own problems before the weekly review |
Most companies in 2026 sit at level 1 or 2. The jump to level 3 is not a better model. It is shared context. An agent at level 2 knows one team's tools. An agent at level 3 knows the company.
What a self-driving company needs underneath
A self-driving car needs more than a driver. It needs a map, sensors, rules of the road and a record of every decision. The company version has the same parts.
1. A map: one current picture of the company
Every agent needs to know what is true now: the current owner of an account, the decision that replaced last quarter's plan, the commitment made on yesterday's call. If each agent rebuilds that picture from scratch on every task, they disagree with each other and repeat work. The map has to be shared, kept current as facts change, and carry the source of every fact.
2. Sensors across systems and conversations
The official state of a company lives in systems of record: ERP, CRM, HRIS, BI. The live state lives in email, chat and meetings, where people say the shipment will slip before the ERP shows it. A self-driving company reads both and uses one to check the other. An agent that sees only one person's inbox sees a road with most of the signs missing.
3. Rules of the road: permissions that follow every source
"Whatever credentials someone handed them" is the failure mode. Agents should see exactly what the person they act for may see, inherited from each source system, with a failed permission check treated as no access. Access rules belong to the company, not to whoever set up the agent.
4. A named driver for every decision
Masad's people choose the destination. In practice that means a named person approves before anything reaches a customer, a supplier or a colleague, and every action records who approved it and why. That answers "nobody can see what they did."
5. A test track
Car makers test autonomy in simulated worlds before real roads. A self-driving company deserves the same: a realistic company with known problems, scored by someone other than the vendor.
How do you test a self-driving company? Simulation
The Glass Company is that test track. Era by Eon built Wisteria, a simulated payments company with 1,200 people, 193 paying customers and fourteen systems, where every dashboard is green. Eon then hid 497 problems inside it and told nobody where.
Sentra read the company for 102 minutes and found 495 of the 497 without being asked a single question. Claude Opus, given the same systems, found 8%; with a checklist of what to look for, 24%. Sentra had never seen Wisteria before the test, and Eon built and scored it, not Sentra. On a second company, it found 500 of 501.
One of the problems shows why the map matters. A $775,000 customer called Doric Worldwide looked healthy in Salesforce. Finance had cut its last renewal by $770,000 with no reason on record, support had promised a fix that failed 29 more times, engineering had logged two incidents against it, three sales reps were chasing it at once, and its close date had moved 19 times. Each team saw one piece. The account owner's own AI assistant, which could read only her inbox, said everything was on track. Sentra joined the five pieces and raised it to her, and a person approved the message before it reached her.
You can explore Wisteria yourself, without signing up, at /glass-company.
Where Sentra fits
Sentra is the map and the judgment under a self-driving company. It does not replace the agents a company already runs. It gives them the same picture.
- It reads how the company runs. Email, chat, meetings and documents, joined with ERP, CRM, HRIS and BI.
- It keeps what is true now. Every fact carries its source and when it became true, so a changed decision replaces the old one instead of competing with it.
- It acts first. It finds what is going wrong, works out who should act, and tells them, before the weekly review.
- A named person approves. Nothing is sent until the right person says yes, and every answer respects each source's existing permissions.
- Every agent reads the same memory. ChatGPT, Claude, Copilot, Gemini and internal agents connect over MCP and REST.
At Lenskart, a top product line showed about 95% in stock in reports while shelves held it at 12.8% availability, and about $860K of sales were lost in the two weeks before the fix. With Sentra across 3,200 stores, time to root cause went from 3 to 4 days to about 30 minutes. The full story is in the Lenskart case study.
How ready is your company? A ten-question checklist
Score one point for each yes.
1. Could an agent tell you, right now, who owns your ten largest customer accounts?
2. Do your agents see the same version of a decision after it changes?
3. Does every fact an agent uses carry its source?
4. Do agents inherit each system's permissions, rather than one shared service account?
5. Can you see every action an agent took last week, and who approved it?
6. Do agents read conversations as well as systems of record?
7. Does anything raise a problem nobody asked about, before the weekly review?
8. Does each raised problem go to a named owner?
9. Does a person approve before anything reaches a customer or supplier?
10. Have you tested your agents on a company where the problems are known in advance?
Zero to three is level 1. Four to six is level 2. Seven or more and you are building a self-driving company. If you want to see what level 3 looks like on your own data, book 30 minutes with us.
Frequently Asked Questions
What is a self-driving company?
A self-driving company is an organization where AI agents carry out most of the execution, such as investigations, reviews, triage and analysis, while people set direction, make tradeoffs and own outcomes. Amjad Masad of Replit introduced the term in July 2026.
Does a self-driving company mean no employees?
No. Masad's own definition keeps people in charge of the destination, the tradeoffs and accountability. The change is that people direct the work instead of performing every step.
What is the difference between a self-driving company and an autonomous company?
The terms are used almost interchangeably. "Autonomous company" is older and sometimes implies agents running a business end to end. "Self-driving company" stresses that people still choose where the company goes, the way a passenger chooses the destination.
What do AI agents need to run a self-driving company?
They need one current picture of the company that every agent shares, signals from both systems of record and conversations, permissions inherited from each source, a named person approving consequential actions, and a record of what was done. Without shared context, agents disagree with each other and repeat work.
How do you test whether agents are ready to run parts of a company?
Test them on a realistic company where the problems are known in advance and someone other than the vendor does the scoring. The Glass Company, built and scored by Era by Eon, is one such test: Sentra found 495 of 497 hidden problems, and Claude Opus alone found 8%.
Is Replit's self-driving company the same as Sentra?
No. Replit describes agents doing work across its own teams and offers Replit for Business to build them. Sentra is the company memory those agents read from: what is true now, who owns it and who may see it, with a person approving before anything is sent. They are complementary layers.