# YC's Summer 2026 RFS, read against the chatter

URL: https://www.thedeepfeed.ai/posts/2026-04-30-yc-rfs-summer-2026-decoded/
Category: Business
Published: 2026-04-23
Updated: 2026-05-22
Author: the-deep-feed
Tags: y-combinator, requests-for-startups, agents, mcp, founder-strategy
Kind: deep

> Y Combinator's Summer 2026 Request for Startups thread reached 8.65M impressions in a week. The chatter underneath the list is more useful than the list.

## TL;DR

- The single most-amplified reaction to the launch was not from a YC partner — it was Packy McCormick joking that "in 2026, your dad read the YC Request for Startups and just built one of the ideas on the list." 10,056 likes, 2.96M impressions. That tweet is the entire thesis.
- Software for Agents and Company Brain pulled the most public claimers within hours — and the OSS leaders in those lanes (mem0 at 54k stars, Composio at 28k, Kong at 43k) are already two years ahead of any Demo Day team.
- Tom Blomfield amplified all 14 RFS items in the launch minute and then re-shared Hyperspell founder Conor Brennan-Burke twice the next day — the only sustained partner-amplification signal of the week.
- Sorted by replies-plus-quote-tweets, the most-discussed RFS items were Company Brain (313 R+Q), Software for Agents (111), and Inference Chips for Agent Workflows (102). Sorted by likes, the leaderboard is different — and the founder-engagement signal is the one that matters.
- The silent RFS items — Counter-Swarm Defense, Industrial Capabilities in Space, Supply Chain 2.0 for Semiconductors — are where the partners are openly recruiting and almost no public claimers exist.

The most-amplified reaction to YC's Summer 2026 Request for Startups did not come from a partner. It came from [Packy McCormick](https://x.com/packyM/status/2048864764416192526), the writer behind *Not Boring*, two hours after the list went live:

> Mom how did we get so rich?
>
> Back in 2026, your dad read the YC Request for Startups and just built one of the ideas on the list.

10,056 likes. 2.96 million impressions. By the rough math of launch-week engagement, that single tweet pulled more attention than the 15 RFS items underneath it combined. It is also the single sharpest editorial on the document YC put out — funnier and meaner than anything the chorus of replying founders said about themselves.

That is the artifact worth opening with, because it sets up the actual question of the week. If reading the RFS were a strategy, the joke would not land. The joke lands because everyone in the replies recognised what they were watching: a few hundred founders treating a public partner-wishlist as if it were a build-spec, while a smaller number of partners and operators tried, mostly in vain, to talk them out of it.

This piece is a reading of that chatter — not the list. The list itself runs on the [YC RFS page](https://www.ycombinator.com/rfs), 15 prompts written by six partners. The interesting object is the public market test the list triggered the moment it went up: a launch thread that pulled roughly 8.65 million impressions in a week, several thousand replies and quote-tweets across 15 items, and a set of patterns about who showed up, who didn't, who got amplified by partners, and what the OSS leaderboard already says about the categories.

![Three zones of the 2026 RFS — saturated lanes, contested categories, asymmetric silence](/post-images/yc-rfs-summer-2026-decoded/hero-three-zones.jpg)

The framing that holds up across all of it is simple. Where the loudest founders piled on, the partner reading the application is going to be the most skeptical reader in the room. Where there is silence, there is asymmetric upside. Where there is a 50,000-star OSS repo doing eighty per cent of the job, the odds of out-shipping it on a 12-week sprint are bad regardless of how exciting the partner's tweet was.

# The launch by the numbers

Strip the editorial out and the launch looks like this. The intro tweet pulled 8,930 likes and 4.27 million impressions on its own:

> AI has stopped being a feature and started being the foundation.
>
> We're excited about a new wave of startups rebuilding software, services, and silicon — and pushing AI into the physical world.
>
> — [@ycombinator](https://x.com/ycombinator/status/2048834285197812146), April 27, 2026

The 15 RFS-item tweets together added another 4.4 million impressions. The CTA tweet for the application picked up another sliver. Total launch-thread reach for the week clears 8.65 million impressions, with an aggregate reply-and-quote-tweet count in the low thousands.

Inside that, the engagement is uneven in a way that matters. Likes track *broad sentiment*; replies and quote-tweets track *founder engagement*, the people who felt the need to either claim the lane, push back, or attach their own product to it. The two leaderboards do not look the same.

# The RFS by founder engagement, not by likes

| Rank | RFS item | Partner | Likes | Impressions | R+Q |
|---:|---|---|---:|---:|---:|
| 1 | Company Brain | [Tom Blomfield](https://x.com/t_blom) | 1,400 | 1.13M | 313 |
| 2 | Software for Agents | [Aaron Epstein](https://x.com/aaron_epstein) | 561 | 308k | 111 |
| 3 | Inference Chips for Agent Workflows | [Diana Hu](https://x.com/sdianahu) | 396 | 692k | 102 |
| 4 | AI for Low-Pesticide Agriculture | [Garry Tan](https://x.com/garrytan) | 701 | 390k | 96 |
| 5 | AI-Native Service Companies | [Gustaf Alströmer](https://x.com/gustaf) | 854 | 330k | 95 |
| 6 | Hardware Supply Chain | [Nicolas Dessaigne](https://x.com/dessaigne) | 561 | 487k | 94 |
| 7 | AI Operating System for Companies | [Diana Hu](https://x.com/sdianahu) | 353 | 105k | 66 |
| 8 | Counter-Swarm Defense | [Tyler Bosmeny](https://x.com/bosmeny) | 267 | 162k | 46 |
| 9 | Dynamic Software Interfaces | [Aditya Gupta](https://x.com/agupta) | 298 | 153k | 29 |
| 10 | AI-Personalized Medicine | [Aditya Gupta](https://x.com/agupta) | 252 | 52k | 24 |
| 11 | SaaS Challengers | [Jared Friedman](https://x.com/snowmaker) | 310 | 154k | 23 |
| 12 | Industrial Capabilities in Space | [Adi Oltean](https://x.com/AdiOltean) | 167 | 86k | 18 |
| 13 | Sell to Biggest Companies | [Brad Flora](https://x.com/bradflora) + [Harshita Arora](https://x.com/aroraharshita3) | 236 | 140k | 14 |
| 14 | Electronics in Space | [Philip Johnston](https://x.com/PhilipJohnston) | 179 | 86k | 13 |
| 15 | Supply Chain 2.0 for Semiconductors | [Diana Hu](https://x.com/sdianahu) | 110 | 51k | 7 |

R+Q is the column to read. Likes are a passive endorsement; replies and quote-tweets are someone walking into a public room and saying, in effect, *I am working on this, or I am pushing back on this.* That is the founder population a partner is looking at when she opens the application portal.

The top three (Company Brain, Software for Agents, Inference Chips) are the noisy lanes. The Company Brain figure of 313 R+Q is roughly three times the next item; the partner who wrote it, Tom Blomfield, has the loudest single thread of the launch. The bottom four (Industrial Capabilities in Space, Sell to Biggest Companies, Electronics in Space, Supply Chain 2.0 for Semiconductors) collectively pulled fewer replies than Company Brain did on its own. That is the first rough cut of where partners will be hyper-skeptical and where they are openly recruiting.

It is also worth flagging what the table does not capture. The single highest-engagement post of the launch week, by a wide margin, is the McCormick tweet above. Inside the partner side of the chatter, the top non-RFS-item tweets are [Jared Friedman's *"Make something agents want"*](https://x.com/snowmaker/status/2048839996254416906) (608 likes), [Aaron Epstein's identical line posted independently](https://x.com/aaron_epstein/status/2048846894827507908) (320 likes), and [Tyler Bosmeny's *"There's no shortage of important problems to work on in the AI era"*](https://x.com/bosmeny) (198 likes). The partner-side conversation about the RFS is, in other words, mostly partners talking to each other in a register the chorus of founder-claimers rarely matches.

# What the parody arc tells you

The most useful piece of dark comedy in the launch is the speed at which the *vocabulary* of a YC category formed. Within minutes of the Software for Agents tweet, [Friedman dropped *"Make something agents want"*](https://x.com/snowmaker/status/2048839996254416906), a four-word inversion of Paul Graham's *Make something people want.* It hit 608 likes inside an hour. Twenty-eight minutes later, [Aaron Epstein posted the same line](https://x.com/aaron_epstein/status/2048846894827507908) independently, in his own thread, and pulled another 320. By 19:39 UTC on launch day, [@rajatsandeepsen had inverted it](https://x.com/rajatsandeepsen) into the actual joke: *"make something agents want, but human pays for it."* That post got zero likes and 66 impressions.

The compression is the point. The crystallization happened in under an hour. Two YC partners, working independently, converged on the exact same four-word slogan. A founder watching it from the outside wrote the parody version, which is the *correct* version of the line, because agents do not pay anyone, and pretending otherwise is the failure mode of the entire category. He was rewarded with 66 impressions. By Monday morning, every founder applying to YC under the Software for Agents item was using a variant of *make something agents want*. By Tuesday, every partner reading those applications had become wary of the phrase.

That is what category formation sounds like in real time. It is also a warning. Once a phrase is being parodied within an hour of its first use, the obvious version of the company under it is being built by twenty teams already, and the partner has already started discounting the language. Founders applying to Software for Agents under that banner are not, in any meaningful sense, claiming a wedge — they are claiming the wallpaper.

![Founder mention volume across 15 RFS categories — a long tail with one outlier in Company Brain](/post-images/yc-rfs-summer-2026-decoded/category-bar-chart.jpg)

# Software for Agents — the saturated lane

The agents lane is over. Within forty-eight hours of [Aaron Epstein's RFS tweet](https://x.com/aaron_epstein), enough YC-grade founders had publicly claimed it that you could fill a directory.

[Haakam Aujla](https://x.com/haakamaujla) posted screenshots of [AgentMail](https://x.com/agentmailapi) "circa April 2025" to assert prior art (98 likes / 16k impressions). [Adi Singh](https://x.com/adisingh/status/2048859708090700155), AgentMail's co-founder, replied with the only optimistic line in the chorus worth quoting — *"it is still very very early to build this btw, go build it"* (82 likes). [Wojtek Szkutnik](https://x.com/wojtekszkutnik/status/2048847211937591806) plugged Prelint, a context layer for coding agents that reads meeting notes and tickets and exposes them via CLI and MCP. [Utpal Nadiger](https://x.com/utpalnadiger/status/2048861407798866322) said his team is *"building aws, but built for agents from first principles"* (the product is OpenComputer, 17 likes). [Scott Motte](https://x.com/motdotla), the original author of `dotenv`, posted Vestauth, an auth layer for agent identities (2 likes). [Javi Sánchez](https://x.com/jvrsanch/status/2048872412566143447) plugged Holocron, *"a living AI-Twin of your company"* via Rebolt (15 likes). And within the first three hours, [@DamsHtg replied to the SfA tweet](https://x.com/DamsHtg) with three words: *"This is basically @composio."* Zero likes, zero impressions, the right answer.

The GitHub picture under that chorus is uglier than any single founder is willing to say out loud. The MCP plumbing layer has more than three hundred repositories above five hundred stars. [punkpeye/awesome-mcp-servers](https://github.com/punkpeye/awesome-mcp-servers) sits at 86,092 stars. The reference servers from [Microsoft](https://github.com/microsoft/playwright-mcp) (31,898) and [GitHub](https://github.com/github/github-mcp-server) (29,437) are in production. [Vercel's agent-browser](https://github.com/vercel-labs/agent-browser) is at 31,337. [PrefectHQ/fastmcp](https://github.com/PrefectHQ/fastmcp) is at 24,947. [Composio](https://github.com/ComposioHQ/composio) is at 27,985 with several hundred integrations shipped. [Google's A2A protocol](https://github.com/a2aproject/A2A) is at 23,548. The plumbing is done. The question, and the one [Eric Migicovsky](https://x.com/ericmigi/status/2048874437467107758), the founder of Pebble, kept hammering in the replies, is what is left to build.

His own line, posted to a Hardware Supply Chain reply, applies just as well to Software for Agents: *"why not just go to SZ to build it there? for the same reason the valley is the center of the world for startups, sz is center of the world for hw."* Translate that to agents: the OSS plumbing is the new Shenzhen. Building generic agent tooling against punkpeye and Composio is the equivalent of opening a hardware-prototype shop in Cupertino instead of Bao'an District. It is not impossible. It is just expensive in the worst way — late, contested, and competing against the people who got there first.

The wedges left under SfA are narrow. Programmatic signup and billing for agents has zero serious players. Agent-native pricing primitives sit at maybe a hundred stars across the field. Reputation and safety registries for MCP servers are toy repos. Real wedges, but each one is a single product, not the category. Anyone walking into a partner interview with a generic "tooling for agents" pitch should expect the partner to read the incumbent list back inside the first thirty seconds.

# Company Brain — the contested lane and the Tratar reply

Company Brain pulled the loudest single skeptical tweet of the entire RFS launch, and it pulled it from someone with credibility on the exact problem. [Zach Tratar](https://x.com/zachtratar/status/2048847052306841913), ex-founder of [Embra](https://www.linkedin.com/posts/zachtratar_im-thrilled-to-announce-that-embra-has-been-activity-7422697953544544256-PzDY) (acquired by Notion in 2025) and now working on Notion AI, replied to the Company Brain tweet with three words: *"Have you tried Notion?"* 219 likes. 88,279 impressions. 46 replies. Zero public response from any YC partner.

That tweet is going to live in the memory of every founder pitching this category in a partner interview, partly because Tratar's bona fides on the problem are real. Embra was, by his own description, an attempt to build exactly the *executable, agent-shaped knowledge layer* the Company Brain RFS now describes, and it ended in a Notion acquisition. That is, in turn, the tightest possible answer to *why has nobody built this yet*. People built it. The platform absorbed them.

It is not a cheap shot. [Tom Blomfield's framing of Company Brain](https://x.com/t_blom) (*"every company has critical know-how scattered across people's heads, old Slack threads, support tickets"*) describes a problem that Notion AI, Glean, Slack AI, Dust and at least a dozen vertical Q&A bots have been chasing for two years. And the YC pipeline is already crowded with claimants. Hyperspell is the most prominent — a YC F25 company explicitly building "Memory & Context Layer for AI Agents", whose founder [Conor Brennan-Burke](https://x.com/contextconor) was [retweeted by Tom Blomfield twice](https://x.com/t_blom) on April 28, the day after launch. [Ontora](https://x.com/ontoratech), in the current YC P26 batch, posted *"Imagine you could talk to every employee in your company within 24 hours."* Javi Sánchez at Rebolt has been shipping Holocron for over a year.

The Hyperspell pattern is the only sustained partner-amplification signal of the entire week. Tom Blomfield amplified all 14 RFS-item tweets inside the [18:39–18:40 UTC minute on April 27](https://x.com/t_blom), the most active partner amplification of the launch, and then re-shared Brennan-Burke's Hyperspell threads twice the next day. That is a different kind of public commitment than a generic retweet of a partner's RFS, and any founder reading the Company Brain RFS as a fresh greenfield should read the Hyperspell pattern as a clear partner anointment of the YC company already inside the lane.

GitHub has the rest of the picture. [mem0ai/mem0](https://github.com/mem0ai/mem0) is at 54,570 stars as the universal memory layer. [letta-ai/letta](https://github.com/letta-ai/letta) is at 22,405. [topoteretes/cognee](https://github.com/topoteretes/cognee) is at 16,972. [Tencent/WeKnora](https://github.com/Tencent/WeKnora) is at 14,156 and ships document-to-RAG-to-reasoning-to-wiki end-to-end. The personal-memory layer is a solved OSS problem for individuals. The exact thing Tom Blomfield's RFS describes, a tribal-knowledge compiler that ingests Slack and Linear and emits an executable skill file, has [nex-crm/nex-as-a-skill](https://github.com/nex-crm/nex-as-a-skill) at 40 stars and a literal repo titled *Sankeerth28/company-brain-agentic-ai* at 2 stars. Nobody serious has shipped it.

So the category is not dead. It is contested in a specific way. The horizontal Q&A bot is dead because Notion exists, and Tratar is the public reminder. The wedge that survives is vertical — Company Brain for legal firms with regulated-data ingestion, or for hospitals with HIPAA-aware permissions, or one that emits *skills* instead of returning text. A founder who walks into a partner interview with the horizontal pitch is going to get the Tratar question read back to them. A founder who walks in with a permissioned, regulated-data, vertically-scoped Company Brain that survives Tratar is in the strongest applicant slot of the batch.

# Inference Chips for Agent Workflows — the under-discussed item

![Capital-intensity moat — three founder figures held at distance from a vertical silicon stack](/post-images/yc-rfs-summer-2026-decoded/inference-chips-moat.jpg)

The third-most-discussed RFS item by R+Q is also the one the public conversation has largely missed. [Diana Hu's Inference Chips tweet](https://x.com/sdianahu) pulled 692,526 impressions and 102 replies-and-quotes (more founder engagement than anything other than Company Brain and Software for Agents) but the engagement is structurally different. There are almost no claimers. The replies are mostly questions, recruiter-style outreach, and amateur takes on Modular's MAX, vLLM, and Tenstorrent's open repos.

That is informative. Inference Chips is the only RFS item where high engagement coexists with low public claiming. The reason is capital intensity. Building inference silicon is a $50M-Series-A business at minimum, and the YC-stage founder pool with relevant experience is small enough that almost everyone capable of applying is doing so privately. The OSS landscape underneath the item is also genuinely incomplete. [Modular's MAX](https://github.com/modular/modular) (open-sourced under Apache 2.0 in early 2025) is the closest thing to a portable inference runtime that runs across NVIDIA, AMD and custom silicon. [vLLM](https://github.com/vllm-project/vllm) at 30k+ stars owns the LLM-serving layer. [Tenstorrent's tt-metal and tt-buda](https://github.com/tenstorrent) repos are public but small. There is no analogue to mem0 or Composio at the inference-runtime layer that a YC team could simply out-ship.

That gap is what Hu's tweet is calling toward. The wedge under it is not "build a chip" — it is "build the agent-workload inference primitive that does not yet have a 50k-star OSS leader." Speculative execution for tool-calling, batched inference for multi-agent orchestration, on-device fallback for sensitive workloads. The category is genuinely under-built, and the chatter is genuinely thin. By the engagement-density rule, that combination is one of the highest-signal asymmetries in the launch week.

# Counter-Swarm and the silent items

[Tyler Bosmeny's Counter-Swarm Defense item](https://x.com/bosmeny) pulled 267 likes and 46 R+Q, mid-pack engagement and not the void the casual reader of the launch thread might assume. He also posted [a separate cross-cutting tweet](https://x.com/bosmeny/status/2048842209571774526), *"If you want to help make America safer, I want to meet to you,"* which pulled another 111 likes and 38k impressions. A third tweet, *"There's no shortage of important problems to work on in the AI era,"* hit 198 likes. That is the most active partner-side recruitment posture of the launch.

The structural point about defense, space and semiconductor RFS items is that the public-claimer signal is muted by design. Defense founders run stealth. Space-manufacturing founders are mostly off X. Semiconductor founders are off X *and* off LinkedIn for non-trivial reasons. So the noise floor under those items is not because the categories are dead — it is because the founder population has incentives to be quiet.

That makes those four items (Counter-Swarm Defense, Industrial Capabilities in Space, Electronics in Space, Supply Chain 2.0 for Semiconductors) the asymmetric pile. Combined R+Q across all four is 84, less than a quarter of Company Brain on its own. Combined application volume into YC for those four is, almost certainly, lower by an even wider margin. If you have the background, the math is straightforward: lower competition for the slot, partner posting publicly that they want more applicants, and an OSS ceiling that does not exist for hardware in the way it exists for software.

[Garry Tan's only RFS-touching tweet of the week](https://x.com/garrytan/status/2048903788891926661), posted that evening, frames the same point from the other side: *"AI helping people looks like this. If we do some of these things in YC's recently released RFS's re housing, food and transportation…"* That is a partner publicly tying his own RFS for low-pesticide agriculture into a broader social-impact framing — and it is the only message-board-style RFS endorsement Garry posted all week. The signal is that the partners have an argument they are tired of making, which is that the asymmetric items on the list are the ones that matter. The chorus is not listening, which is exactly what makes those items asymmetric.

# The wedge taxonomy

![Four quadrants of the 2026 RFS — saturated, contested, asymmetric, niche — with the two red dots marking where the upside actually lives](/post-images/yc-rfs-summer-2026-decoded/wedge-taxonomy-quadrants.jpg)

Crossing the founder-engagement signal with the OSS-incumbent signal gives a sharper four-quadrant read of the entire RFS than reading either column on its own. The vertical axis is *founder R+Q* on the launch tweet, the public-claimer count. The horizontal axis is *closest OSS leader's GitHub stars*, the open-source ceiling.

**High engagement, high OSS — the death lane.** Software for Agents and Company Brain both fall here. Company Brain has 313 R+Q and a 54k-star OSS leader in mem0. Software for Agents has 111 R+Q and an 86k-star OSS leader in awesome-mcp-servers. A team applying with the horizontal version of the pitch is competing against ten public claimers and an OSS project two years ahead. The only viable wedge is a vertical or a regulated-data integration the OSS will never ship.

**High engagement, low OSS — the race.** AI-Native Service Companies and AI for Low-Pesticide Agriculture sit here. Loud public claim-rates, but no real OSS contender at scale. The category is open, but the application pool is stacked with founders pitching the obvious version. The wedge is operational: regulatory licenses, last-mile field deployment, an actual book of business by Demo Day. The applicant who wins is the one who arrives with the part the OSS cannot build because it is not a software problem.

**Low engagement, low OSS — the asymmetric upside.** Counter-Swarm Defense, Industrial Capabilities in Space, Electronics in Space, Supply Chain 2.0 for Semiconductors. Few public claimers, no OSS ceiling, partners actively recruiting. This is where the math favors the founder if she has the credentials. The reason the chatter is sparse is the reason the upside is real.

**Low engagement, high OSS — the niche.** SaaS Challengers (CRM-shaped variants) and parts of Dynamic Software Interfaces fall here. Public claim-rate is modest, but the OSS leaders ([Kong at 43k](https://github.com/Kong/kong), [CopilotKit at 30k](https://github.com/CopilotKit/CopilotKit), [ag-ui at 13k](https://github.com/ag-ui-protocol/ag-ui)) own the horizontal layer. The wedge here is narrow, vertical, and frequently dependent on a single big customer.

The four-quadrant frame is a better cut than the standard *saturated / contested / asymmetric* division because it forces the founder to read both signals at once. Most categories on the 2026 RFS that look interesting in isolation collapse on closer reading once the OSS column is filled in.

![Founder density per RFS category — clustering, contested zones, and the empty squares](/post-images/yc-rfs-summer-2026-decoded/founder-density-heatmap.jpg)

# The OSS kill list

For each RFS item that touches software, the question worth asking before opening the application is: how many stars does the closest OSS competitor have, and how mature is it. Use this as a check before you write the wedge.

| RFS item | Closest OSS leader | Stars | What it kills |
|---|---|---:|---|
| Software for Agents | [punkpeye/awesome-mcp-servers](https://github.com/punkpeye/awesome-mcp-servers), [Composio](https://github.com/ComposioHQ/composio) | 86,092 / 27,985 | Generic MCP tooling, agent integration hubs |
| Company Brain (memory) | [mem0](https://github.com/mem0ai/mem0), [Letta](https://github.com/letta-ai/letta) | 54,570 / 22,405 | Personal memory layers, "remember the agent's conversations" |
| Dynamic Software Interfaces | [CopilotKit](https://github.com/CopilotKit/CopilotKit), [ag-ui](https://github.com/ag-ui-protocol/ag-ui) | 30,576 / 13,000 | Generative UI libraries, chat-shaped UI kits |
| AI Operating System for Companies | [Tencent/WeKnora](https://github.com/Tencent/WeKnora), [Portkey gateway](https://github.com/Portkey-AI/gateway) | 14,156 / 11,000 | Document-RAG, AI gateways with guardrails |
| Sell to Biggest Companies | [Kong](https://github.com/Kong/kong), [agent-governance-toolkit](https://github.com/microsoft/agent-governance-toolkit) | 43,000 / 1,300 | API/AI gateway category, agent governance toolkits |
| SaaS Challengers (CRM) | [relaticle](https://github.com/relaticle/relaticle), [open-pencil](https://github.com/open-pencil/open-pencil) | 1,300 / 4,500 | OSS AI-native CRM and design tooling |
| AI-Native Service Companies | None at scale | <100 | Tools exist; the outcome-delivery business is empty |

The simple rule: if the closest open-source competitor sits above twenty thousand stars and has been in development for more than two years, the wedge has to be a vertical (legal, healthcare, manufacturing, defense), a regulated-data integration the OSS cannot ship, or a primitive missing from the existing stack. *"Better mem0"* is not a wedge. *"mem0 for hospitals with audit-grade access logs"* is.

The other lesson from the GitHub data is when the OSS leader is the *enabler* of the RFS rather than its competitor. Kong at 43k stars is bad news if you want to build "an AI gateway." It is good news if you build something that lives downstream of Kong and ships only the thing Kong cannot do — usually because the thing Kong cannot do involves a regulator, a license, or a contract that an open-source project cannot sign.

![A 12-week YC sprint plotted against multi-year head starts of the OSS incumbents](/post-images/yc-rfs-summer-2026-decoded/cohort-timeline.jpg)

The cohort-math is brutal. A YC team has roughly twelve weeks from kickoff to Demo Day to ship something an OSS project with three years of momentum cannot. Paul Graham's old framing, that the best ideas are the ones nobody has gotten around to or cannot bring themselves to do, keeps surfacing in the replies. The first pile is what the GitHub table on this page is full of. The second pile is the asymmetric pile. Vertical compliance, regulated data, liability: the parts no OSS maintainer wants to own.

# AI-Native Service Companies — the contrarian view

The most damaging tweet of the launch *for the AI-Native Service Companies item* did not come from a venture critic; there were no high-signal venture critics in the public thread that week. It came from [Bill Zahalan](https://x.com/bilalzahalan/status/2048844498596171917), who wrote: *"feels like the mobile era framing a bit. in mobile, native gave you years. here incumbents are adding AI features in quarters. distribution still wins."* Five likes, 5,300 impressions: small reach, sharp argument.

GitHub agrees with his read. There is no AI-native service business on GitHub, only tools and skills. Bookkeeping has [norman-finance/norman-mcp-server](https://github.com/norman-finance/norman-mcp-server) at 44 stars. Tax filing has [shinkoku](https://github.com/kazukinagata/shinkoku) at 339 for the Japanese e-Tax market. Insurance brokerage tops out at flashvenom/quickfire at 45 stars and that is a human AMS shell. Prior authorization is dominated by Microsoft accelerators in the ten-to-twenty-five star range. Vertical compliance MCPs (SOC2, HIPAA, PCI-DSS, GDPR) sit at zero or near-zero. There is no Stripe-equivalent SDK for shipping AI services, no liability-bearing platform that actually files the tax return, no F100 pilot-in-a-box.

That gap is the asymmetric bet under [Gustaf Alströmer's RFS](https://x.com/gustaf). The service tools exist as OSS. The outcome-delivery business does not. A vertical AI accountant that owns the audit risk; an AI insurance broker that binds carrier policies through carrier APIs; an AI prior-auth engine that signs the SOC2 attestation. Each is a single hard company. Each is a wedge a generalist competitor cannot copy without absorbing the same compliance debt. Bill Zahalan's distribution argument applies to the *tools-shaped* version of the category. It does not apply to the *license-shaped* version, which is the one the partner is actually asking for.

# Which founders the partners retweeted — and ignored

The partner-side amplification pattern is the second signal embedded in the launch thread, and it is simpler than it looks once you stop asking who *should* have retweeted whom and just count what happened.

[Tom Blomfield](https://x.com/t_blom) was the most active amplifier of the launch by a wide margin. He retweeted all 14 of the RFS-item tweets inside the launch minute on April 27. The next day, he retweeted [Conor Brennan-Burke](https://x.com/contextconor)'s Hyperspell thread twice. That is the only sustained, named-founder amplification of any RFS-aligned company any partner ran in the launch week. The signal is unambiguous: Tom Blomfield is openly anointing Hyperspell as the YC reference implementation of Company Brain.

[Diana Hu](https://x.com/sdianahu) was the quietest of the partners-with-RFS-items. She tweeted [a single deadline reminder](https://x.com/sdianahu/status/2048836052413251909) (176 likes), then went silent for the rest of the launch week. She authored three of the 15 items (Inference Chips, AI Operating System, and Supply Chain 2.0) but did not amplify any specific founder under any of them. The reading is not that the partner is uninterested; it is that the founders worth funding under those items are not the ones tweeting at her.

[Garry Tan](https://x.com/garrytan) posted exactly [one RFS-aligned message](https://x.com/garrytan/status/2048903788891926661) — the housing/food/transportation framing tweet. He did not retweet any individual founder.

[Tyler Bosmeny](https://x.com/bosmeny) posted three RFS-recruitment tweets across the week and amplified a small number of replies, none of which got more than ten public likes. That is a recruitment posture, not a public-anointment posture. The defense-tech founders being read by partners are being read in DMs, not on the timeline.

[Jared Friedman](https://x.com/snowmaker) wrote the catchphrase and then refused to elevate any single founder claiming the agents lane. The most-funded software-for-agents team out of the S26 cohort will, by inference from this pattern, be one nobody in the public chorus has heard of yet.

The pattern across all of it is consistent: partners amplified specific founders in *one* category (Tom Blomfield on Hyperspell) and stayed silent everywhere else. The silence is not a void. It is an editorial choice. Founders who got no retweet from the relevant partner during launch week should read that as information, not as the absence of it.

# What got cut from the list, 2024 to 2026

The 2026 RFS is also worth reading against the [archive of past RFS posts](https://www.ycombinator.com/past-rfs/). What got *removed* between 2024 and 2026 says as much about YC's current investment posture as what is on the new list.

**Foundation models.** Off the list for two cycles running. YC has accepted that frontier-model training is not a YC-shaped bet, no matter how cheap synthetic data gets. The closest item on the 2026 list is Inference Chips, which is a hardware-flavored substitution rather than a return.

**Generic crypto.** Off the list. The September 2025 [YC × Coinbase RFS](https://www.ycombinator.com/blog/build-onchain) was the last partner-blessed crypto wishlist; the Summer 2026 RFS proper has zero crypto items. Where the 2024 list had Stablecoin Infrastructure as a hedge category, the 2026 list does not even gesture at it. The implication is that crypto-platform pitches will be read in 2026 the way crypto-platform pitches were read in 2023 — respectful, brief, and a no.

**Generic vertical SaaS.** Off the list. The 2024 RFS had explicit slots for "AI for ⟨industry⟩." The 2026 list folded all of that into SaaS Challengers, which is partner code for *we will fund a wedge into Salesforce, Workday, or Adobe, not a generic AI-for-X.* The chatter under SaaS Challengers is dominated by founders who already know which incumbent they are aiming at.

**General-purpose robotics.** No general-purpose robotics RFS. The 2026 list has Counter-Swarm, Industrial Capabilities, and Hardware Supply Chain — all robotics-adjacent, all framed as missions. Founders who tweeted *"I'm building a humanoid robot company"* under the launch thread got the deafening silence that platform-shaped pitches now reliably draw.

**Climate.** Almost off the list. AI for Low-Pesticide Agriculture is the only clean climate item on the 2026 RFS. The 2022 [Climate Tech RFS](https://www.ycombinator.com/blog/rfs-climatetech/) listed *over a hundred climate startups together worth more than $10B*; the 2024 RFS had three explicit climate items; the 2026 RFS has half of one. That is a downward step almost nobody on X commented on — and it is the loudest single signal in the document about what YC has decided is no longer a YC-shaped thesis at the partner-meeting table.

**Consumer.** Off the list completely. Not one consumer-app RFS in 2026. The 2018 RFS had nine items including consumer-leaning categories; the 2024 list had explicit consumer slots; the 2026 list has zero. The Dynamic Software Interfaces item is sometimes read as a backdoor consumer slot, but [Brad Flora](https://x.com/bradflora) and [Harshita Arora](https://x.com/aroraharshita3) framed it explicitly as enterprise-app rendering, and the partner replies bear that out.

The combined message of those omissions is sharper than any single RFS item on the page. YC has decided, in public, that 2026 is not the year for foundation models, generic crypto, generic vertical SaaS, general-purpose robotics, climate, or consumer. The list is what is left after all of that has been crossed out.

# The historical hit rate

![Past RFS cohorts plotted as bars across years — most produce a long tail; one rare red bar marks the unicorn-magnitude breakout](/post-images/yc-rfs-summer-2026-decoded/historical-hit-rate.jpg)

The standard sanity check on any RFS reading is the historical one. YC has been publishing public RFS posts since 2008 in various forms, and the modern public-RFS cadence stabilised around 2018. The empirical pattern is well-known among the partners and frequently elided in the founder chatter.

The strongest counter-example is at the top of the canon. *Stripe* (YC S09) was not on any YC RFS; there was no public "online payments" item the year Patrick and John Collison applied. *Airbnb* (W09) was not on any RFS. *Coinbase* (S12) was not on a public RFS list. *DoorDash* (S13) is the cleanest case where an RFS post and a unicorn intersect: YC had published interest in restaurant and local-commerce ideas around the time DoorDash applied, and Tony Xu's [demo-day pitch](https://www.ycombinator.com/blog/doordasy-yc-s13-raises-2-dollars-dot-4m-to-improve-food-delivery-growing-over-20-percent-weekly/) leaned into that frame. *OpenAI* was never on an RFS, and was never a YC company in the standard batch sense; it began as YC Research, a separate entity that pre-dated the modern RFS cadence.

The pattern on the explicit RFS items is mixed. Sam Altman's 2014 [Breakthrough Technologies RFS](https://blog.samaltman.com/new-rfs-breakthrough-technologies) listed energy, AI, biotech, transportation, and education — categories that produced [Boom](https://boomsupersonic.com/), [Helion](https://www.helionenergy.com/) (YC W14), and [Astranis](https://www.astranis.com/) (YC W16) over the following decade, but no Stripe-magnitude breakouts. The 2018 [Updated RFS](https://www.ycombinator.com/blog/updated-request-for-startups) added longevity, government, programming tools, transportation, and underrepresented-founders categories; the longevity slot produced [Loyal](https://loyalfordogs.com/) and a handful of biotech-adjacent companies, the government slot produced one small cluster, and the rest produced the long tail of forgettable applications you would expect. The 2022 [Climate Tech RFS](https://www.ycombinator.com/blog/rfs-climatetech/) coincided with a YC climate-portfolio that scaled to $10B in aggregate value but had a small number of outsized winners and a much larger number of write-downs. The 2024 [Updated RFS](https://www.ycombinator.com/blog/ycs-latest-request-for-startups) by Dalton Caldwell explicitly named LLM-application categories that produced the bulk of the 2024–25 batch but, again, no obvious unicorn-sized hits this early.

The pattern across all of it is that a public RFS is a weak predictor of a single category's eventual unicorn output. Some of the largest YC outcomes were in lanes the partners had not articulated. Some of the most heavily-RFS-anointed lanes underperformed against the median batch. The signal in any given RFS is *which categories the partners want more applicants in*, which is useful for the founder calibrating their approach — but it is not a forecast of which categories the partners will eventually fund the breakout in.

For 2026, the implication is that reading the chatter is the more useful exercise than reading the list. The list tells you what the partners are excited about. The chatter tells you what the *applicant population* looks like — and that is the variable that determines whether a given founder's wedge clears the bar in the partner interview.

![Addressable surface area surviving the 2024–2026 commoditization wave](/post-images/yc-rfs-summer-2026-decoded/tam-shrinkage.jpg)

# A pre-mortem for the S26 cohort

The chatter contains enough information to predict the failure modes already.

The most-funded category will have the highest mortality. Software for Agents will likely take the largest single share of the S26 cohort and the highest 18-month attrition rate of any category in the batch. The reason is structural: ten teams in the same category facing the same OSS incumbents and the same five enterprise buyers. Two find a vertical wedge and survive. Two get acquired by mid-tier infra companies. The remaining six spend 2027 looking for a pivot.

The contested category will produce one breakout and seven graveyards. Company Brain has the same dynamic, but worse, because the OSS competitors there (mem0, Letta, Cognee) are fast-moving, well-funded, and shipping. The breakout will almost certainly be the vertical play, in legal, healthcare or financial services, with regulated-data ingestion as the wedge. Anyone trying to ship "Notion AI but better" will be obsolete by Demo Day.

The asymmetric categories will produce zero or one company each, and one of them will be the best company in the batch. Counter-Swarm Defense, Industrial Capabilities in Space, Supply Chain 2.0 for Semiconductors, and AI for Low-Pesticide Agriculture each have effectively zero public claimers. YC will fund one or two teams across all four. Statistically, given how thin the application pool is and how aggressively partners are recruiting, the surviving teams in those categories will have a much higher probability of becoming top-quartile of the batch than the saturated-category teams will.

The AI-Native Service Companies item will produce one breakout and a lot of $1M-ARR plateaus. The exit from the plateau is owning a regulated-license stack: the actual CPA, the actual broker, the actual AMS. Any team that walks in with *"we're an AI service"* and walks out without *"we own the license"* is going to discover the Series B problem the hard way.

The pre-mortem is mostly bad news for the loudest applicants. That is the point. Reading the chatter is supposed to be discouraging in the specific sense that it discourages the founder from applying with the version of their idea that everyone else is also applying with.

# How to read an RFS

The decision rules that fall out of the chatter:

1. Count public claimants in the first seventy-two hours. If more than five recognisable founders or YC alumni publicly claim the lane, assume the partner has read all of those tweets. Walk in with a wedge that is not on any of those landing pages.
2. Find the loudest skeptic and absorb the question. *"Have you tried Notion?"* is not an obstacle to getting funded — it is the question you need a clean answer to. Track the most-engaged skeptical reply on each RFS item and bake the response into the pitch.
3. Check the OSS ceiling. If the closest open-source competitor has more than twenty thousand stars and is two years old, the default position is that you are building a vertical, a regulated-data integration, or a primitive on top of it. If you cannot articulate which, do not apply with that pitch.
4. Notice the silence. RFS items with low reply-and-quote counts and high-credibility partner authorship (Bosmeny on Counter-Swarm, Oltean on Industrial Capabilities, Johnston on Electronics in Space, Dessaigne on Hardware Supply Chain) are the asymmetric ones. The chatter density is the application density. Less competition for the slot is the gift the founders ignoring those items are handing you.
5. Read the partner amplifications. Tom Blomfield retweeted Hyperspell's founder twice in the 24 hours after launch. That is the only specific named-founder anointment of the week. Tailor the application accordingly. If the partner has already publicly anointed a YC company in the lane, the pitch needs to be the *complement* to that company, not the alternative to it.

The framework is opinionated on purpose. The point is not to talk a founder out of a category — it is to make sure that whichever category she walks into, she has an answer for the version of the partner who has been reading the same X thread she has.

# The list is the prompt; the chatter is the answer key

Y Combinator publishes an RFS for the obvious reason: it tells founders what partners are excited about and gives them cover to apply outside the obvious lanes. It also functions as a marketing event, and the marketing event generates a public market test that did not exist when YC started doing RFS posts in 2008.

That market test is the actual product. The 8.65 million impressions across the launch thread told several thousand founders, for free, which categories already have eight visible founders building them and which categories have zero. GitHub told them, also for free, which categories have a 50,000-star OSS reference and which categories have a two-star toy repo. Reading both is half an hour of work. Skipping it and applying with the generic "Software for Agents" pitch is the version of the application that ends up in the no pile.

The list is the prompt. The chatter is the answer key. Packy McCormick's joke is the warning label on the cover.

# Update, May 22, 2026 — three weeks in

Two updates since this piece shipped. First, on May 20, OpenAI offered $2M of API credits to every company in YC's Spring 2026 and Summer 2026 batches in exchange for equity, with the instrument an uncapped SAFE that converts at the next priced round. YC moved the Summer application deadline by 48 hours so more founders could qualify. The full deal mechanics, including the Sam Altman, Tyler Bosmeny, Jared Friedman, and Rohan Varma quotes, are in [Sam Altman writes a $2M token cheque to every YC company](https://www.thedeepfeed.ai/posts/2026-05-22-openai-2m-tokens-yc-batch-equity/).

Second, the Spring 2026 batch has now launched 55 of its 171 companies publicly, which gives a first empirical read on the predictions above. The Software for Agents prediction held; 16 of 55 visible companies sit in that one lane. Three of the four asymmetric silent lanes (Industrial Space, Supply Chain 2.0, Low-Pesticide Agriculture) confirmed at zero claimers. Counter-Swarm Defense broke at exactly one (Arlo Industries, the passive aerial sensing mesh play). The biggest broken prediction is the horizontal Company Brain call. Four visible Spring 2026 companies pitch horizontal company-knowledge layers, and YC funded all four. The full three-week field check, with the per-lane scorecard and six companies worth watching, is in [Three weeks after the YC Summer 2026 RFS, the Spring batch is the field check](https://www.thedeepfeed.ai/posts/2026-05-22-yc-spring-2026-rfs-three-week-field-check/).

## Sources

- [Y Combinator — Requests for Startups](https://www.ycombinator.com/rfs)
- [YC RFS launch thread (April 27, 2026)](https://x.com/ycombinator/status/2048834285197812146)
- [Packy McCormick — "Mom how did we get so rich?"](https://x.com/packyM/status/2048864764416192526)
- [awesome-mcp-servers (GitHub)](https://github.com/punkpeye/awesome-mcp-servers)
- [mem0 — universal memory for agents (GitHub)](https://github.com/mem0ai/mem0)
- [CopilotKit (GitHub)](https://github.com/CopilotKit/CopilotKit)
- [Kong API + AI Gateway (GitHub)](https://github.com/Kong/kong)
- [Composio (GitHub)](https://github.com/ComposioHQ/composio)
- [Letta — stateful agents (GitHub)](https://github.com/letta-ai/letta)
- [Hyperspell (YC F25)](https://www.ycombinator.com/companies/hyperspell)
- [YC past Requests for Startups archive](https://www.ycombinator.com/past-rfs/)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-04-30-yc-rfs-summer-2026-decoded/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt