Y Combinator observation: turn public signals into constrained experiments

What this dated observation is for

This note records public YC material read on 2026-10-04. It is neither an endorsement nor evidence that a company, product, or RFS heading is suitable for a particular team. YC Requests for Startups says it offers examples of what YC would like to see, rather than a set of required applicant projects. Treat a public signal as a prompt to name the friction, assumptions, and smallest safe experiment.

1. 1YC public material
2. 2state the user friction and unverified assumption

1. 1One constrained prototype
2. 2measure authority, data, and failure boundaries

1. 1Evidence
2. 2continue, revise, or stop

Read RFS as problem framing, not a market forecast

The 2026 Fall RFS page was the material observed on that date. Three headings provide useful translations into work.

A Cloud for Small Software

A Cloud for Small Software describes the difficulty of sharing and deploying purpose-built software for small groups. Its hard parts include authentication, permissions, and safely sharing arbitrary code. Build a two-person CRUD tool and test invitations, permission revocation, non-exposure of secrets, audit logging, and deletion. Measuring deployment speed alone misses the central sharing boundary.

Multiplayer AI

Multiplayer AI frames a long-running agent that several people can observe, redirect, and hand off. Limit an experiment to one read-only research task. Define shared visibility, who may change instructions, the change history, conflict priority, and stopping. Evaluate answer quality separately from the team’s authority to make a decision; shared access without a control boundary creates ambiguity rather than collaboration.

Self-Maintaining APIs

Self-Maintaining APIs imagines a provider detecting client code affected by a breaking change and creating repair pull requests. Use a public sample repository and one synthetic breaking change. Measure detection recall, unintended-edit rate, pull-request diff readability, test pass rate, and the invariant that nothing merges without human approval. Permissions and secret boundaries for an external agent are as material as transformation logic.

Read companies through integration friction

Use the YC Startup Directory to discover a company, then return to that company's documentation, security material, pricing, terms, and published source for product claims. The following pages were read as public descriptions at the observation date. Their batch or active status belongs to the YC company page; product features belong to their product documentation. Neither is a benchmark or procurement recommendation.

Airbyte’s YC page identifies the company as active W20. Its documentation describes data replication and a context layer for AI agents using MCP, a Python SDK, and an HTTP API. The friction is connecting, synchronizing, and searching scattered business data. Connect only two dummy SaaS sources in a small experiment. Record sync latency, schema change handling, credential rotation, access control, and deletion propagation. A large connector catalogue does not guarantee source quality or data governance.

PostHog’s YC page describes analytics, session replay, feature flags, experimentation, LLM observability, and more on one product surface; see its documentation. The friction is operating separate observation tools to determine whether a shipped feature was used and whether a change caused an outcome. Use anonymized staging events, two flag cohorts, a preregistered primary metric, exposure events, and a kill switch. An SDK does not establish causality when bot traffic, small samples, or concurrent releases remain.

Supabase’s YC page identifies it as active S20 and lists Postgres, Auth, Row Level Security, Realtime, and APIs. The documentation covers database, auth, storage, realtime, edge functions, and framework guides. Its friction is the initial assembly of backend foundations. Use fictional data with two organizations and two roles, and automatically test that RLS rejects cross-tenant access through API, direct SQL, Realtime, and Storage. Fast setup does not remove responsibility for RLS policy, migrations, backups, region choice, and operations.

Preserve the boundary on a four-hour refresh

A future four-hour refresh should retain RFS batch names and headings as dated page observations. Keep YC company batch/status separate from documentation feature claims and separately record version or release information. Try one candidate at a time in an isolated environment, retain success and failure conditions, withhold secrets, and record deletion steps. The value of the observation is a smaller hypothesis that can be tested, not a bundled trend claim.

Follow-up: the rendered directory is a discovery surface, not a static dataset

On 2026-10-04 JST, the official Startup Directory rendered in a browser with its title, batch/industry/region filters, and a loading company list. The visible page says YC has invested in over 5,000 companies; this is YC's display, not an independently audited market statistic. Its dynamic rendering means this observation does not claim a complete directory export, a current company count, or a verified update date.

The 2024–2026 pattern is to separate company discovery from product verification. After a directory or RFS lead, record the visible company URL and batch/status as a dated YC observation, then independently inspect the company's documentation, security, pricing, terms, and release notes. Design one integration with dummy data and a removal procedure. A company page or RFS topic can justify a hypothesis; it cannot prove demand, reliability, funding health, customer adoption, or suitability.

A dated 2024 baseline for reading RFS

YC’s February 16, 2024 early-interview notice states that a new RFS list had appeared earlier that week and clarifies that applications were not limited to those themes. This is a dated historical signal of what YC wanted to encourage, not a count of companies created or a current application deadline. Do not infer an exact RFS publication date from “earlier this week.”

For a longitudinal exercise, record the text and date of one request, the later company’s own description, and the customer problem it actually addresses. Keep “theme encouraged,” “company listed,” and “working product independently verified” as separate observations. Compare the 2024 source with the directory observed October 4, 2026; a mutable company profile cannot establish when a feature shipped. A request that later attracts many listings may still leave the original user problem unsolved. Read customer workflow and adoption evidence before inferring that a startup category has matured.