Andreessen Horowitz runs an early-stage program called speedrun. It writes first checks into new companies, hosts them for twelve weeks in San Francisco and ends with a Demo Day. Because it sees thousands of applications per batch, its partners get an early view of what founders are building, and they write about it in public.
This article reads that public material: the program's own pages, its newsletter, its partners' requests for startups and a16z's wider writing on AI-native companies. The ideas that recur point toward the same place as Y Combinator's recent requests: companies that notice, decide and act with far fewer people coordinating by hand. Here is what speedrun says, what it implies for companies that are not startups, and where the gaps are.
Key takeaways
- speedrun invests up to $1M per company and runs a twelve-week residency in San Francisco.
- Its partners' 2026 requests center on proactive agents, tools for the "100x solo founder" and products built for agents.
- a16z partners argue that coordination costs set company size, and that AI is starting to break that link.
- For established companies, the lesson is to move from asking AI questions to having a system that observes and proposes action.
What is a16z speedrun?
a16z speedrun is an accelerator run by Andreessen Horowitz for founders at the earliest stage. The terms for the current cohort are public. According to a16z's SR008 application post, the firm invests "up to $1M: $500K for 10% upfront via a SAFE," plus up to another $500K in the company's next round within 18 months. Founders also get "$10M+ in cloud, AI, and software credits from 200+ partners," and a16z says "we don't take board seats." The SR008 residency runs from 1 February to 18 April 2027, in person at a16z's San Francisco office, with Demo Day on 13 April.
The program's homepage puts its philosophy in one line attributed to Marc Andreessen and Ben Horowitz: "We invest in the magnitude of a strength, not in a lack of weaknesses." It continues: "At the start, the strength is the founders and whatever they know that nobody else does."
Cohorts are numbered. SR004 ran in early 2025, SR005 in Los Angeles later that year, SR006 and SR007 in 2026, and SR008 starts in February 2027. Each batch publishes requests and lessons through a weekly newsletter, which is the best public record of what the partners are looking for.
Two features of the program shape what its founders build. The first is speed: a two-week close and a twelve-week residency push teams to ship a working product quickly, which favors small teams that lean on AI for engineering and operations. The second is selection on founders over categories, which means the companies reflect what strong founders choose to build when AI is cheap and capable, more than a thesis set in advance by the firm.
How much of speedrun is AI?
speedrun did not begin as an AI program, and its general partner Andrew Chen still describes it as broad. In a May 2026 newsletter post answering applicants' questions, he wrote that it is "a horizontal program," with B2B, B2C, infrastructure, healthcare, gaming, govtech and robotics teams "alongside the AI applications." He added: "We care about the founders more than the category." He also mentioned that the previous batch had 70 companies.
Even so, the share of AI companies is high. When applications opened for SR005 in April 2025, PocketGamer.biz reported that SR004 had 42 companies selected from more than 6,000 applicants, and that 69% of them were AI-native. It quoted Josh Lu, then a principal at a16z: "We're backing some of the most ambitious entrepreneurs building AI-first products, next-gen entertainment and entirely new interactive experiences."
The practical reading: speedrun selects founders first, and the founders it selects overwhelmingly build with AI at the core. Their products are a fair sample of what AI-native companies look like in their first year.
What do speedrun partners say they want to fund?
In May 2026, as SR007 applications closed, the speedrun newsletter published a set of requests from its investors. Several of them, read together, describe how an autonomous company would work.
| Request | Partner | Key line |
|---|---|---|
| Proactive AI agents | Kenan Saleh | Shift from 'ask to answer' to 'observe to act' |
| Tools for the 100x solo founder | Emily Bennett | "Scale meant headcount. AI is starting to break that relationship." |
| Agent-native products are coming | Fareed Mosavat | "The real opportunity is products designed for agents from scratch." |
| GUIs for agents | Jonathan Lai | "bring the power of agents to normies everywhere" |
| World of infinite output | Lejla Johnsen | "systems that can make decision-making visible, trackable, and improvable" |
| Come for the agent, stay for the network | Troy Kirwin | An agentic platform can aggregate demand and flip the power balance in offline markets |
Kenan Saleh's note is the shortest and the most important for operators: "Today's AI products are reactive. You give the model a prompt, it responds with an answer." The shift he wants is to products that observe what is happening and act. That is the difference this site draws between reactive and proactive systems in reactive, proactive and predictive intelligence.
Emily Bennett's note explains why companies could get much smaller. "For decades, company size has been dictated by coordination costs," she writes. "More customers meant more support reps. More revenue meant more finance and ops." Her request is for "startups building tools that allow a single founder to operate what would previously have required entire teams."
Three themes, one direction: the autonomous company
Strip out the startup framing and the speedrun requests describe three properties of a company that runs more of itself.
- It observes. Work starts when the system sees something, such as a delayed shipment, a promised quote or a slipping deadline, and not only when someone types a prompt.
- It coordinates without headcount. Bennett's point about coordination costs is the economic core. If a system can route work, chase follow-ups and assemble status, the company no longer needs a person for each of those jobs.
- It acts through agents. Mosavat and Lai both assume agents will do real work and need software built for them. A company that can delegate a task to an agent, and check the result, can grow output without growing the team at the same rate.
Lejla Johnsen's request supplies the fourth property, the one most often missing: decision-making that is "visible, trackable, and improvable." A company that observes and acts but cannot show who decided what, and whether it worked, is harder to trust than one that acts slowly. This is why the levels of company autonomy are better climbed one at a time.
What does a16z say more widely about AI-native companies?
speedrun sits inside a larger firm, and two pieces of a16z writing fill in the picture.
The first is from a16z's annual Big Ideas 2026 list, published in December 2025. Sarah Wang, a general partner on the Growth team, predicted that "the system of record will finally start to lose primacy." Her reasoning: models can now read, write and reason across operational data, turning records systems "from passive databases into autonomous workflow engines." In her view, "the interface becomes a dynamic agent layer." For an operating company, that suggests value moves from where data is stored to the layer that decides what to do with it. The same list included speedrun partners' ideas, including Emily Bennett's picture of an AI-native university in which "schedules optimize themselves."
The second is Andrew Chen's October 2026 essay "New Founder Archetypes in the AI Age". It is careful about limits. "Coding might get easier while sales, recruiting, judgment, and persistence don't," he writes. And: "A solo founder might be able to do the work of ten people, but a team of three might be able to do the work of thirty." On solo founders, he asks whether "someone still needs to do a million meetings with customers."
That last point matters. The speedrun view is optimistic about small teams, and it is also clear that judgment and relationships stay human. The site's article on the one-person unicorn covers those limits in numbers.
What speedrun's ideas mean for established companies
Most readers run companies that already exist, with customers, staff and systems. The speedrun ideas still apply. They translate into concrete moves.
| speedrun idea | What it means inside an existing company | First step |
|---|---|---|
| Observe to act | A system watches email, calendar and chat for situations that need action | Connect one team's sources read-only |
| Coordination without headcount | Status, follow-ups and handoffs are assembled by software, not meetings | Replace one weekly status meeting with an owned list |
| Agent-native work | Routine tasks are delegated to agents with clear inputs and outputs | Pick one task, such as drafting follow-ups, and hold it for approval |
| Visible decisions | Every action has an owner, a date and a record of the outcome | Require proof before an item counts as done |
The order matters. Observing comes before acting, because a system that acts on a wrong picture of the company is worse than none. Visible decisions come before wider delegation, because the record is what tells you which tasks can safely run with less oversight. For the delegation step itself, see AgentOps on running many agents without losing control.
To know whether the shift is happening, track a few plain numbers each month. They show whether the company is still asking or has started observing.
| Measure | Asking company | Observing company |
|---|---|---|
| Who starts the work | A person notices and asks | Most items are raised by the system from sources |
| Time from signal to owner | Days, often found at the weekly meeting | Hours, with one owner named |
| Status meetings per team per week | Several | One or none, replaced by an owned list |
| Items closed on proof by their date | Not measured | Counted every week, with misses explained |
| Share of routine tasks delegated | Close to zero | Rising, each one approved until the record holds |
The last row should move slowly. A company that delegates one new kind of task a month, and only after the record shows the previous one holds up, ends the year far ahead of one that tried to automate everything in a quarter and pulled it back after the first public mistake.
Common mistakes when copying AI-native startups
- Copying team size targets. A seed-stage software company and a manufacturer with 300 staff have different work. Copy the principle, fewer people coordinating by hand, and leave the headcount numbers.
- Buying reactive tools and calling it proactive. A chat assistant waits for a question. If nothing reaches a person unless they ask, the company has not moved to "observe to act."
- Letting agents act before the picture is right. An agent that sends a follow-up for an order already cancelled costs trust. Start with proposals a person approves.
- Ignoring the human bottlenecks. Chen's list of sales, recruiting, judgment and persistence applies to every company. Free people's time for those, and measure whether it happens.
- No record of outcomes. Without a log of what was done and whether it worked, the company cannot learn which work to delegate next.
How GeniOS approaches the autonomous company
GeniOS is built around the "observe to act" shift. It reads the places where a company's work happens, such as email, calendar, chat and the business tools it already uses, read-only to start. It forms the company's goals from its own emails, decks and meetings, and checks every team's pending work against them.
What reaches a person is intelligence: one situation, what to do, why, one owner and a date. The owner answers Delegate, Done, Later or Not relevant. Delegate hands the work to a person or an AI agent, and nothing is sent until a human approves. An item closes only on proof, such as the email that went out.
Each outcome feeds back, so the system learns what this company treats as important. That is the "visible, trackable, and improvable" decision-making speedrun's partners ask for, applied inside an existing company, on the way to the autonomous company.
Worked example: from asking to observing at Vaultex
The names, dates and numbers in this example are made up to explain the idea. Vaultex is a safety-footwear maker in Greater Noida and GeniOS's first design partner.
In the reactive version of the week, Harsh in sales opens a chat assistant on Wednesday and asks it to draft a quote for Bata: 12,000 pairs of S3 safety shoes at $12 a pair, about $144,000. The draft is fluent. It does not know that the steel toe cap supplier raised prices 8% on Monday, that Bata wants 60-day payment terms, or that Line 2 is busy with Reebok samples. Harsh sends it. The vendor review is on Friday.
In the "observe to act" version, nobody has to ask. The system reads the same sources and raises three items on Monday afternoon:
- For Harsh: the Bata quote is due before Friday's review and is priced on the old toe-cap cost. Revise the margin by Wednesday. An AI agent prepares the revised quote, and Harsh approves before it goes.
- For Anushka in production: Line 2 is booked for Reebok samples. Confirm a slot for the Bata order by Thursday so the quote can promise a date.
- For the finance owner: Bata expects 60-day terms on about $144,000. Decide before Friday.
Meanwhile a fourth item sits in Anushka's list for next week: Adidas has a vendor audit coming, and the IS 15298 test reports need to be gathered. Each item closes when its proof exists: the sent quote, the slot in the plan, the written decision on terms, the attached reports. One founder or a small team can run this, because the coordination is done by the system. The Vaultex case study shows the full picture.
Conclusion
a16z speedrun's partners describe companies that observe and act, coordinate without adding headcount, delegate to agents and keep decisions visible. Those four properties define an autonomous company as well as any formal definition. Established companies can adopt them in order: observe first, then propose, then delegate with approval, then widen what runs on its own as the record proves it safe. To see what that looks like with your company's data, see pricing or book a call.
Sources
- a16z speedrun, Program homepage, accessed November 2026
- a16z, Applications for speedrun SR008 are now open, 14 October 2026
- a16z speedrun newsletter, questions re: a16z speedrun (Andrew Chen), 14 May 2026
- a16z speedrun newsletter, That's a wrap on SR007 applications, 19 May 2026
- PocketGamer.biz, a16z speedrun opens applications for fifth event with up to $1m on offer, 2 April 2025
- a16z, Big Ideas 2026: Part 1, 9 December 2025
- a16z, New Founder Archetypes in the AI Age (Andrew Chen), 30 October 2026