Andreessen Horowitz has 1,100 portfolio companies tracked on the network. 379 of them publish APIs. Of those, 12 reach the agent-native band, the top tier of Agent Readiness. The firm’s highest-rated API provider, OpenAI at 86.0, is not among the twelve.
When this series last looked at a16z, in July, the network counted 186 publishing companies and 8 exemplars. It now counts 379 and 19. The portfolio did not double in ten weeks. The network’s view of it did, so read the growth as our coverage catching up, not a16z changing.
The twelve
| Company | Agent Readiness | Kin Score |
|---|---|---|
| Mem | 61.3 | 42.5 |
| Stripe | 58.3 | 81.6 |
| Wefunder | 55.9 | 50.7 |
| Dwolla | 52.9 | 61.6 |
| Ripple | 51.8 | 55.6 |
| Svix | 49.4 | 76.7 |
| Anchorage Digital | 47.5 | 47.1 |
| ModernFi | 46.4 | 42.4 |
| Lightspark | 46.2 | 66.4 |
| Nash | 42.1 | 61.4 |
| Blackbird | 40.4 | 43.4 |
| Method | 39.4 | 67.3 |
Eight of the twelve move, hold or raise money: Stripe, Dwolla, Ripple, Anchorage, ModernFi, Lightspark, Method and Wefunder. Svix delivers webhooks, and Nash dispatches deliveries. Mem, an AI note-taking workspace, leads the twelve at 61.3, with a Kin Score of only 42.5. Agent Readiness is not a ranking on its own: Zapier, at 64.7, scores higher than any of the twelve and is still gated at agent-ready.
The frontier labs sit a band lower
a16z’s AI bets lead on Kin Score: OpenAI at 86.0 and ElevenLabs at 83.8, both exemplar. Both are agent-ready, at 48.1 and 50.9, and both score high enough for agent-native. The gate holds them back. Agent-native requires idempotency and error semantics. Both labs have verified error semantics, and neither documents idempotency. The companies building agents do not yet publish the one rail an agent needs in order to retry a write safely against their own APIs. Payments companies have shipped idempotency keys for a decade, because a retried request that charges a card twice is a real loss.
That is the pattern in the table. The agent-native band rewards the discipline money forces on an API: safe retries and explicit errors. Fintech learned it from duplicate charges. AI labs have not yet had to.
The rest of the book
Across the 379 publishers the composite average is 38.1, thin, and Agent Readiness averages 23.7, agent-aware. The bands run 19 exemplar, 40 strong, 115 developing, 121 thin, 71 emerging and 13 minimal. Across all 1,022 scored portfolio companies, publishing or not, the composite average falls to 18.3. Well-known names stay closed or early at the network layer: Anduril at 36.2 and Shield AI at 13.4. Safe Superintelligence and Character.AI publish nothing.
What would move it
For the portfolio, the lever is narrow and cheap. Of the 124 portfolio entries at agent-ready, 45 already score high enough for agent-native and are held back by the gate, and 97 lack documented idempotency. A documented idempotency key would move more of this book than any new API. The fintech companies in the portfolio already show how.
See the banded portfolio at apis.io/vcs/a16z/.