The Last Three Batches

The Last Three Batches

Y Combinator publishes a Request for Startups: a list of the things its partners most want someone to build. It is the single most-read document in early-stage tech, and it is a genuinely useful artifact — a well-informed group saying out loud what they think the next decade needs.

So we ran it as a benchmark.

We took the RFS themes in force across Winter 2026, Spring 2026 and Summer 2026, the companies YC actually funded against them, and audited all of it against what already runs on this platform. Thirty-nine asks. Then — because an audit you grade yourself is worthless — we sent 62 adversarial agents at our own answers under three separate hostile lenses, each instructed to refute, each told to default to refuted when uncertain.

They killed 46 of our 62 claims.

That number is the point of this piece. Everything below survived a process designed to destroy it.

The scoreboard

Across 39 asks spanning three batches:

Eleven of thirty-nine, built without a batch, without a partner, without a seed round. And six of the thirty-nine we will simply never do, which we will say plainly rather than stretch a definition to cover.

What YC funded, and what already runs

Infrastructure for AI workloads

YC funded The Token Company — token compression via a drop-in API.

We run MicroScheduler, an inference orchestration layer with fair-queue scheduling across backends, plus KV-cache optimisation and kernel-level tuning. And the compression piece is not a slide, it is aither-kvcache, MIT, on PyPI: TurboQuant at 2–4 bits, 3.8–7.1× against fp16, provably within 2.7× of the information-theoretic optimum, no calibration dataset, no offline pass. A 70B-class model at 128K context drops from 40 GB of KV cache to 10 GB.

It implements Zandieh et al., arXiv:2504.19874. You can read the method, check the maths, install it this afternoon.

Verdict: SHIPPED. Survived all three lenses.

DevTools for AI agents

YC funded Polymath (agent workflow rehearsal), Salus (guardrail validation) and Corelayer (on-call engineers) — three companies, three slices of one problem.

The refuters were handed our claim that the Workforce agent system covers the whole lifecycle and told to tear it apart. The verdict came back:

The Workforce agent system is a complete, implemented, and deployable platform. All claimed components — 9 multi-agent packs, skill.md playbooks, budget enforcement, approval workflows, MCP tool routing, and safety policies — are present.

Markdown-driven agent specs, capability-based tool routing, budget governance, approval gates. The rehearsal, the guardrails and the on-call loop are one system here because they were always one system.

Verdict: SHIPPED.

A Cloud for Small Software

YC funded InsForge — agent-native AWS for full-stack development.

We ship AitherZero for the bootstrap layer, a tenant-app spine, portal-kit for the UI, and a provisioning gateway that mints tenant-scoped, expiring, resource-bound keys. adk up gets you a running agent. Invoke-AitherPlaybook node-onboard turns a bare machine into a secure cluster node on Windows, Linux or macOS from one command.

Verdict: SHIPPED.

Company Brain

YC funded Luel — multimodal training-data curation.

We run AitherGraph: code understanding with a real call graph, RAG, event graphs and memory graphs unified into one queryable plane, self-hostable, with cross-domain search. An agent here does not grep. It asks where is auth enforced and gets the symbol, its signature, its callers and its callees.

Verdict: SHIPPED.

Compliance

YC funded Fenrock AI — bank back-office automation.

Our GDPR claim went to the refuters and came back verified with something better than a code read:

GDPRCompliance.py contains fully implemented, executable code — not a design doc, checklist, or stub. It orchestrates cascade deletion across 8 subsystems (Identity, Secrets, Strata, Chronicle, TenantDB, TenantWorkspace, Consent…)

and, from a second lens:

fully implemented and live via public API endpoints (tested: curl http://localhost:8001/gdpr/retention-policy returned 200)

An agent hit the live endpoint rather than taking the file's word for it. That is the standard.

Verdict: PARTIAL — GDPR ships; automated regulatory filing and cross-jurisdiction monitoring do not.

Founder-led SMB software, and consumer AI for a billion people

YC funded Tensol (AI employees with email and phone access) and Reframe (procurement agents); Tsenta (job applications) and Pax Historia (an AI sandbox game with the best retention among LLM games).

We ship aither-adk — pip install aither-adk, twelve providers behind one interface, uncapped by default, bring your own key — plus AitherShell, multi-channel ingest, Bonsai models that run in the browser with no server at all, and an on-prem USB distribution for people who will never have a datacentre.

Verdict: SHIPPED.

Now the part that makes this worth reading

An audit that only lists wins is marketing. Here is what 46 refutations found, published because a claim nobody tried to kill is not a finding.

AitherCare is a design, not a product. All three lenses refuted it independently. The telephony layer is real and shipped — 292 lines, in and outbound, verified. Everything the PRD promised on top of it, the caregiver coordination, the companion-memory binding, the medication tracking, does not exist. One refuter put it exactly right: "Claiming a whole product on the strength of one primitive." Correct. We were.

Multiplayer agent sessions are not built. No session attach or join endpoint, no live turn-stream broadcast, no driver-role model, no handoff. YC funded Lance here, running agents across 50+ hotels. We have the messaging fabric and not the product.

Our agent-to-agent federation is real code bound to localhost. AitherA2A is 3,000+ lines implementing Google's A2A protocol with genuine task delegation and cross-agent session memory — and it listens on 127.0.0.1. A refuter called it: "A shipped product must be publicly accessible." It is internal infrastructure. We will stop describing it as anything else.

Capture is design-only. Academy is twelve markdown documents. The Primer — the ask about teaching children to read and reason — is the one we would most like to answer and the one where we have written the most and built the least.

And our skills are static markdown files. We claimed dynamic MCP-driven discovery as a point of difference against gstack's hardcoded slash commands. Two of three lenses refuted it: aither-skills installs .md files into fixed paths, which is functionally what gstack does. The MCP gateway serving 1,272 tools is a genuinely different mechanism, but it is not what the skills pack is. We were describing the platform and crediting the repo.

That correction matters for what comes next, so hold onto it.

The audit found we were wrong about ourselves — in the other direction too

Our own prior internal audit, from 2026-07-28, recorded that WebAuthn attestation was "entirely NOT FOUND" and filed proof-of-humanity as a critical gap.

All three lenses refuted that. WebAuthnManager.py is 640 lines, implements full FIDO2 attestation verification, and is wired into the identity service with nine production endpoints for registration and authentication. It has been there the whole time.

We had been under-reporting a shipped capability for five days because one audit said so and nobody re-checked. A wrong claim in your own favour costs you credibility; a wrong claim against yourself costs you a feature. The refuters caught both directions, which is the entire argument for running them.

What this says about the other side of the table

Now the uncomfortable part, and we will keep it to what is checkable.

gstack. Garry Tan — YC's CEO — shipped an open-source Claude Code framework that reached 89,700 GitHub stars in under two months. It packages his development philosophy as slash commands: /office-hours, /plan-ceo-review, persona skills that play CEO, Eng Manager, Staff Engineer.

The adoption data from its own power users, after four weeks: 6 of 35 commands survive into stable daily use. Seventeen percent. The design and CEO-mode commands land flat on teams. The one component everybody keeps is /qa — which drives a real browser, takes screenshots, and verifies the UI actually rendered.

Look at what that is. /qa is a tool with an execution loop and a check that can fail. It touches reality. It is the one part of the box that does not work by pretending to be a person, and it is the one part that survives. The framework ran the experiment and published the result: personas get forked out, tools survive.

Eighty-nine thousand stars measure distribution, not product. Garry Tan has one of the largest founder audiences on earth attached to the most powerful brand in startups. The stars prove the machine works. They say nothing about whether persona prompts improve software, and the field data says mostly not.

QM. YC open-sourced what it called a multi-agent harness. Version 0.1.0. Its architecture is per-scope isolation with no agent-to-agent delegation — all interaction runs through a human. "Multiplayer" means several people talking to the same bot. Its own threat model declares it unsuitable for external customer access, and it documents six production gaps.

We are not mocking a v0.1.0 for being a v0.1.0. Shipping early is correct. What is worth naming is the gap between what the artifact does and what the megaphone said it was, and the fact that a hundred thousand engineers now believe a single-agent Slack bot is what multi-agent architecture looks like.

That is the cost of distribution without a check that can fail. It is the same failure mode as a persona reviewing its own work: the loop closes, everyone feels good, and nothing was tested.

The honest comparison

We are not claiming we would beat those companies. Several are better at their one thing than we are, and Pocket shipping 30,000 units at 50% month-over-month growth is a harder, realer achievement than any architecture diagram. Execution against a market is the game. We have not played it at that scale.

What the audit actually shows is narrower and stranger:

Eleven of thirty-nine RFS asks are already answered by one integrated system. Not eleven companies. One platform — 262 services, 196 containers, an MCP gateway serving 1,272 tools, 22 quality gates in CI — where the compliance layer, the inference scheduler, the agent runtime, the knowledge graph and the identity plane are the same system and know about each other.

YC's model funds the slices. It is a portfolio strategy and it works: you back twenty narrow companies, most fail, a few compound. Nothing here argues against it.

But it does produce a particular blindness. When you fund The Token Company for token compression, Polymath for agent rehearsal, Salus for guardrails and Corelayer for on-call, you have funded four companies that will each build a quarter of an agent platform and then spend years integrating with each other. The RFS is written by people who can see that the pieces belong together — that is why they are on one list — and the funding vehicle can only buy them one at a time.

The integrated version is not more clever. It is just cheaper to build once than to reassemble from four cap tables.

Why we ran 62 agents at our own work

Because the alternative is what everyone else does.

Every claim in this article carries a file path an agent opened, a live endpoint that returned 200, or a source we fetched. Where the evidence did not hold, the claim is not softened — it is gone, and its corpse is in the section above.

That is the same discipline that runs everywhere else here: every number on our public repo pages is generated from the code being shipped and asserted afterwards by a gate with a self-test that proves it can still fail. Twenty-two of those gates run in CI. They gate our marketing copy at the same standard as production code.

It is not a hard bar. It is just one almost nobody clears, because clearing it means publishing the 46 things that did not survive.

The short version

YC's Request for Startups is a good document. Its authors can see that agent infrastructure, compliance automation, inference economics and knowledge graphs are one problem — they wrote them on one list.

Eleven of those thirty-nine asks already run here, as one system, verified by adversaries. Fourteen more are half-built and we have said exactly which half. Five are gaps. Six we will never touch.

Build the whole thing. It is cheaper than you think, and the integration is the product.

---

Every verdict here comes from an agent that read the source, hit the endpoint, or fetched the citation. Batch rosters and RFS themes were gathered from ycombinator.com and public announcements; where a roster could not be confirmed, the row says so rather than guessing. The four repos are open: ADK · skills · kvcache · AitherZero.