OCTOBER 2026 · ABOUT 7 MINUTES
To scale self-improving AI startups, build the grader first
Self-improving AI products grow only as fast as they can tell good work from bad. Five well-funded startups show it: one has passed $1B in run-rate revenue, while the rest are still turning research into products. Below: the sales plays that are working, the moats that will hold, and what each is likely to ship next.
Jump to the six proven sales plays →$1M → $1B
Cognition's run-rate, September 2024 to September 2026
~89%
of Cognition's own code is now written by its agent, Devin
$410M
Recursive's compute deal with AWS, most of its $650M round
400+
companies in Twin1's sales pipeline, at a $20M seed
01 · THE IDEA
These products scale as fast as they can be graded
Code either passes its tests or it doesn't. A benchmark score goes up or down. That's most of the reason the one company here past $1B in revenue sells to software teams, and why Devin now writes about 89% of Cognition's own code. Legal advice, a sales pitch, a strategy memo — nothing grades those automatically. A model can't improve from feedback it never gets — and a product that doesn't improve with use is just selling the same thing to more people, not scaling. Sakana's Darwin Gödel Machine shows the clean version: an agent edits its own code and keeps only the changes that score better on benchmarks.
Hover a dot for examples, or switch to my 2028 view (live site)
Where each dot sits is my judgment, not measured data.
MY TAKE
People talk about self-improvement as if it lives in the model. I've come to think it lives in the data pipeline. If a product never learns whether anyone acted on what it produced, "learning" is just guessing with more steps.
So the feature I'd fight for in any product in this space is a boring one: capture every accept, edit and reject on purpose, from the first day.
02 · WHY THE GRADER COMES FIRST
A better grader is a growth lever
Here's the chain I'd put on a whiteboard for any team building one of these products. Each link is something a product team can measure, and each one ends in revenue.
1
Grade more work, more accurately
Every task gets a reliable good-or-bad signal
2
The product improves with use
More of its work is accepted without edits
3
Customers hand over more
They trust it with bigger tasks and less review
4
Accounts expand
More usage, more seats, more teams per customer
5
The grade becomes the sales pitch
Measured results shorten the next deal, then the loop repeats
On the live site: drag the grader-accuracy slider and watch the curve move
Same product, same customers, three graders: share of work accepted without edits
Illustrative model, not company data. Each month, quality closes 15% of its remaining gap to perfect, scaled by how much better than a coin flip the grader is: q(next) = q + 0.15 × (2 × accuracy − 1) × (1 − q), starting at 50%. A wrong grade teaches the wrong lesson, so a 60% grader is barely better than none.
| Grader level | What grades the work | Revenue it unlocks |
|---|---|---|
| 0 · None | Nobody, systematically | Pilots that stall |
| 1 · Thumbs | Occasional user ratings | Renewals on goodwill |
| 2 · Behavior | Every accept, edit, reject, captured by design | Seat expansion |
| 3 · Automated | Tests, benchmarks, AI judges checked against experts | Usage-based growth, more autonomy |
| 4 · Outcomes | The customer's own business result | Pricing on results; enterprise-wide deals |
Where the five sit today, in my reading. Cognition and Sakana's coding work start at level 3 — tests and benchmarks grade for free, and Cognition's run-rate went from $1M to $1B in two years. Recursive is building its own level 3 for research. Twin1 needs to get from 2 to 4: a partner-hours number its customers will sign. Adaption's real-time learning depends on level 2 signals from production.
What I'd track weekly: how often the grader agrees with expert reviewers, the share of tasks that get graded at all, the false-accept rate (bad work marked good), and the accepted-without-edits rate it should be moving.
03 · THE PLAYBOOKPROVEN SALES STRATEGIES · IF YOU READ ONE SECTION, READ THIS ONE
Six proven sales plays, and how to execute each one
This is the section I'd bookmark. Every play below has a real customer or company announcement behind it. Most of the hard numbers come from Cognition, because it's the only one of the five far enough along to publish customer results. That makes its playbook the one to study, and a template the others can borrow.
8–12×
faster code migration at Nubank, with migration costs cut 20×
Cognition case study
3–4 hrs
per file migrated, vs 30–40 hours by hand
Cognition, 2025 review
100s → 1,000s
of Devin agents planned at Goldman Sachs, alongside 12,000 developers
Goldman CIO, July 2025
+30%
combined enterprise revenue in the 7 weeks after buying Windsurf
Cognition, Sep 2025
<5%
customer overlap with Windsurf, so the deal was almost all new logos
Cognition, 2025
Vendor-reported or executive estimates, not independent audits. Buyers will discount them, which is a lesson in itself: third-party verification is the next proof worth building.
Where each play fits in the sale
PLAY 1 · QUALIFY + LAND
Public case with numbers
Start with the “migration” of your market
The real case. Nubank had an 8-year-old ETL monolith and a migration scoped at 1,000+ engineers moving about 100,000 data class implementations over 18 months. Devin did it 8–12× faster at a 20th of the cost. A second Latin American bank finished a tax-ID system migration with half the planned effort.
Why it works. The work is already budgeted, everyone hates doing it, it splits into thousands of small pieces, and every piece can be checked. That's a painful problem, a measurable result and a built-in grader in one project.
Run it:
- List 20 accounts sitting on a large, repetitive backlog.
- Qualify tasks with Cognition's rule of thumb: if a person can finish one in about three hours, the agent can usually take it.
- Sell a paid slice of the backlog, priced per completed unit.
WATCH
Cost per completed unit vs. the customer's own costTRAP
A flashy showcase project with no clear definition of donePLAY 2 · PROVE
Customer executive, on record
Let the customer's executive tell the story
The real case. Goldman Sachs became the first major bank to adopt Devin. Its CIO, Marco Argenti, went on CNBC calling it “our new employee,” said the bank would start with hundreds of agents and might go to thousands, and estimated a 3–4× productivity gain over earlier AI tools.
Why it works. When a respected buyer says it publicly, every peer's risk drops. The “new employee” framing also moves the conversation from software budget to workforce budget, which is far bigger.
Run it:
- Give one lighthouse account early access and co-design the rollout with them.
- Agree on public storytelling rights in the contract, not after.
- Frame the product as a teammate the customer manages, with human review built in.
WATCH
Inbound from the lighthouse's peers in the 90 days afterTRAP
Going public before results are in, so the story outruns deliveryPLAY 3 · DELIVER
Public case with numbers
Put engineers in the field, then productize what they learn
The real case. Cognition runs a forward-deployed engineering team because generic agent deployments get poor results. At Nubank, the bank's own engineers supplied examples of past manual migrations to teach Devin the pattern, and it got faster and more reliable every day.
Why it works. The first deployments teach you what the product is missing. Field engineers find it; the product team turns it into features the next customer gets for free.
Run it:
- Staff one field engineer per account for the first five customers.
- Teach the agent on the customer's own past examples.
- Turn every repeated field fix into a product ticket within a week.
WATCH
Days from signed contract to first accepted taskTRAP
Field work that never becomes product, which turns you into a services firmPLAY 4 · REFERENCE
Company announcements
Win one industry, then sell to its neighbors
The real case. Devin now works at thousands of companies, with a banking cluster of Goldman Sachs, Santander and Nubank where each name helps sell the next. Sakana is doing the same in Japanese finance with MUFG and Daiwa. Twin1 started in law and is moving to finance and energy, the industries law firms serve.
Why it works. Buyers in one industry share compliance rules, vocabulary and fears. Solve them once and the second and third deals close faster.
Run it:
- Win three logos in one industry before going broad.
- Package what you learned: compliance answers, deployment options, ready-made workflows.
- Ask each customer for one introduction to a peer.
WATCH
Share of new deals that came from a referenceTRAP
Chasing single logos across ten industriesPLAY 5 · PROVE
Company announcements
Publish an honest performance review
The real case. Cognition published a review of Devin's first 18 months: what it's good at, and where it struggles. It said Devin does best with clear requirements and doesn't handle mid-task changes well.
Why it works. Admitting limits builds trust faster than a launch video, and it screens out deals that would have churned. Buyers bring better-scoped projects when they know the boundaries.
Run it:
- Publish a quarterly report: where it works, where it doesn't, with numbers.
- Include a simple rule for which tasks qualify.
- Link it in every proposal.
WATCH
Churn from mis-sold deals; time to qualify a leadTRAP
A report polished by marketing until nobody believes itPLAY 6 · SCALE
Company announcements
Buy distribution, or borrow it
The real case. Cognition bought Windsurf: about $82M in ARR, 350+ enterprise customers, under 5% overlap, and combined enterprise revenue up 30% in seven weeks. Sakana chose Series A investors that doubled as sales channels. Cognition has Microsoft and Infosys ties; Recursive is building infrastructure with AWS.
Why it works. Product quality compounds slowly; distribution can be bought or borrowed in one move. Low overlap is what makes an acquisition grow the business instead of just consolidating it.
Run it:
- Map who already sells to your buyers: IDEs, integrators, clouds, banks.
- Choose investors and partners who open doors, not just write checks.
- If you acquire, check customer overlap first.
WATCH
Net-new logos per dollar spentTRAP
Buying a product you then spend a year mergingThe first 90 days, week by week
MY TAKE
What strikes me about Cognition's playbook is how unglamorous it is. It didn't lead with "AI that writes software." It led with migrations, the work engineers hate, sized in years, where every step can be checked. Then it let Goldman's CIO call Devin a new employee on national TV. Pain, proof, then a public champion — that order matters.
If I were at Twin1 or Sakana, I'd run plays 1 and 2 before anything else: find the migration of their market — the big, repetitive, checkable job a customer already budgets for — win it end to end, and get one respected buyer to say so publicly.
04 · THE FIVE COMPANIES
Moats of the five self-improving AI leaders
| Company | Who uses it, and for what | How it charges | Valuation | Revenue so far |
|---|---|---|---|---|
| Cognition | Software teams clearing backlogs and shipping faster | Not disclosed | $26B (May 2026) | $1B+ run-rate |
| Sakana AI | Japanese banks and brokers building research and advisory tools; researchers using its open tools | Software licensing plus implementation work | ~$2.65B (Nov 2025) | Not disclosed; Daiwa project in production development |
| Twin1 AI | Lawyers, bankers, energy pros answering colleagues without being pulled in | Enterprise SaaS, single-tenant or private cloud | Not disclosed | Law-firm customers; 400+ in pipeline |
| Recursive | Its own researchers, for now, finding better training methods faster | Nothing yet | $4.65B | None yet |
| Adaption Labs | Teams running AI in production who want the same quality for less | Not disclosed | ~$1B (reported) | Not disclosed |
Darker means stronger, in my judgment (click a cell for the reasoning on the live site)
Data only it has
Built into daily work
Trust and permissions
Compute
Cognition
Sakana AI
Twin1 AI
Recursive
Adaption Labs
Twin1 leans on permissions. Six layers of rules and AI checks decide what each twin can share, and an enterprise MCP server lets other agents pull context through those same rules.
Cognition leans on daily habit: an editor, an agent, code review and parallel runs in one place, plus partnerships with Microsoft and Infosys.
Recursive has compute as both its biggest advantage and its biggest bill. The AWS deal took most of the round.
MY TAKE
The defense I think people underrate is permissions. The big labs will ship memory and personalization, no question. What they won't solve quickly is the question every general counsel is going to ask: when an employee leaves, who owns their twin? Whoever answers that clearly — in the product, not in a policy PDF — gets to set the rules everyone else is measured against.
Compute is a different kind of advantage. Very real, but it lasts exactly as long as the money does.
05 · WHAT I EXPECT NEXT
Product roadmap predictions
Each call is grounded in something the company has already said or shipped, with my odds and the signal that will settle it.
| Company | What I expect them to ship | By | My odds | Signal to watch |
|---|---|---|---|---|
| Cognition | Packaged, fixed-price migration products for regulated industries, priced per completed unit instead of per seat Why: It already markets Java 8→17 and .NET 8 upgrades to financial services; migrations have the cleanest grader it owns | Mid 2027 | 65% | A per-unit price on its website |
| Twin1 AI | A self-service version for small firms and individual professionals Why: The company has said it plans to follow the enterprise product with one | Mid 2027 | 60% | A public sign-up page or free tier |
| Sakana AI | A packaged research product for banks, built from Marlin and The AI Scientist Why: Its Daiwa wealth-management work entered production development in August 2026, and it is already turning its research tools into commercial software | End 2027 | 55% | A second bank naming Sakana as its research tool |
| Adaption Labs | Its first commercial product: an API or SDK for models that adapt in production Why: It is hiring designers for interfaces beyond the chat bar, and has a research-automation system in development | Mid 2027 | 50% | Public pricing or a design-partner announcement |
| Recursive | Its research loop opened to outside partners, starting with pre-clinical biology Why: Its CEO has said it plans to expand from AI research to physics, chemistry and pre-clinical biology | End 2027 | 45% | A named partner outside AI research |
| Recursive | The research infrastructure built with AWS offered to other companies Why: AWS and Recursive say the infrastructure may be useful to other companies doing automated research | End 2027 | 40% | A listing or announcement on AWS |
| A frontier lab | Long-term work memory or a “work twin” for enterprises — the move every company above has to plan for Why: Memory and personalization are already shipping to consumers; enterprise is the obvious next step | 2027 | 70% | An enterprise launch with admin and permission controls |
WHAT I'D TELL ANOTHER PRODUCT LEAD
Pick the customer who already keeps score. Fit comes faster when someone can tell you, in numbers, that it worked.
Design the feedback before the model. Every accept, edit and reject is training data, or it's gone.
Find the paying use early. A great team gets you the first round. Usage gets you the second.
Treat permissions as a feature. It's the part the big labs will be slowest to copy.