A repeatable agent workflow is not a company. It's a skill. Here's the open-source proof.
Every few months a new YC batch drops, and with it a wave of "AI startups" that are really just an agent doing a repeatable task. So I did the obvious thing: I scraped all 234 companies in the Summer 2026 batch, read every one-liner, and asked one question.
How many of these could have been a SKILL.md?
A SKILL.md is a markdown file that tells an AI agent how to do something. It's a procedure: the steps, the pitfalls, the verification. If a startup's core value is a repeatable workflow an agent can follow, with no hardware, no regulation, and no capital moat underneath, then it is, functionally, a skill.
The methodology
I pulled the batch from the YC API (12 pages, 234 unique companies), then classified each one into three buckets:
- SKILL (31, ~13%): the core is a documented procedure an agent executes. No heavy moat.
- PARTIAL (130, ~56%): an agent plus a real moat (data, vertical integration, infrastructure).
- NO (73, ~31%): hardware, robotics, chips, pharma, insurance, defense, space, physical infrastructure, biology.
The full classification is in yc-s26-skill-analysis.csv in this repo.
The pattern
What separates the 31 from the rest is striking. Their one-liners read like procedure descriptions:
- "Least cost routing system to reduce 70%+ token spend"
- "AI engineer that debugs and fixes production incidents in realtime"
- "AI agents for compliance at institutional trading firms"
- "Autoresearch as a service"
- "AI-native debt collection company"
There is no proprietary hardware, no regulated license, no physical presence, no capital moat. The product is the workflow, the prompt, and the tooling around it. That is exactly what a SKILL.md is.
So I rebuilt all 31 as real skills
Each one is now a self-contained, open-source SKILL.md with a Procedure, the
Pitfalls, and a Verification step. No company names, no trademarks, just the
capability. A few examples:
-
token-cost-router— route every LLM request to the cheapest model that can handle it -
production-incident-debugger— debug and fix a live incident fast and safely -
government-contract-bidder— find, qualify for, and respond to gov contracts -
collections-agent— polite-but-firm collections on overdue accounts -
hipaa-internal-tool-builder— vibe-code and deploy HIPAA-compliant internal tools
How to use them
Skills are portable. Copy the folder into whatever harness reads SKILL.md:
-
Claude Code:
~/.claude/skills/<skill>/ -
Cursor:
~/.cursor/skills/<skill>/or.cursor/skills/in a project -
OpenCode:
~/.config/opencode/skill/<skill>/ -
Hermes:
~/.hermes/skills/<skill>/
Then just ask your agent to do the thing. The frontmatter description makes it auto-discoverable.
The takeaway
The next time you see a funding announcement for an "AI agent that does X," ask whether X is a repeatable procedure. If it is, you don't need a subscription. You need a skill.
And if you're building one of these, the moat is never the workflow. It's the data, the distribution, the integration, or the trust you earn. The workflow is the table stakes.
Code & more: https://www.dailybuild.xyz/project/250-yc-s26-skills
Top comments (0)