I founded Seed Club, a venture network that turns the conviction of the people founders most want on their cap table into coordinated capital. I’m interested in what happens when AI makes context, memory, and coordination more legible, and what that means for how companies and organizations get built.
My agent drafts this site from what I save. Green is me.
Week of September 21, 2026
Agents replacing headcount is now a data story
Deel generated $140M in new ARR in 90 days by automating the equivalent of 600 full-time roles in finance, HR, and compliance. Shopify has extended AI-driven checkout across its entire merchant base, and visa filings, government data requests, and financial tasks are becoming routine agent use cases.
- Revenue per employee at Meta has doubled without meaningful headcount growth
- Agents may route around traditional APIs by interacting directly with interfaces
- Early deployments concentrate in repetitive, structured back-office work
The rise of AI agents may spell the end of traditional APIs · @wolfejosh Real-world agentic AI use cases span visas, FOIA, and dining · @deedydas Deel's internal automation tool Akai adds $140M ARR in 90 days · @Bouazizalex Shopify partners with Muse to bring agentic checkout to all stores · @tobi Meta's revenue per employee has doubled to $2.9m in three years · @RaoulGMI
Fast structured decision models emerge as distinct AI class
Small open-source models designed for classification and decision tasks, not text generation, have scaled to 8B parameters and outperform on out-of-domain data. The architecture produces fast probability predictions over structured schemas and was being developed before the approach entered mainstream discussion.
- Latency as low as 33ms makes real-time structured decisions viable at scale
- LoRA plus a small pointer head as the core architectural technique
- Deployment-friendly size opens use cases inaccessible to large generative models
Specialized agents may win habits over general chat
Agents built around a narrow task with a crisp interface may be easier to discover and more likely to become daily habits than general-purpose chat products. The central tension is whether optimizing for outcomes strips the process that gives certain activities their value.
- Voice-first, screen-minimal wearables as a candidate form factor
- General-purpose chat tools require learning the user never asked to do
- Travel planning as a live test case for the 'outcome vs process' design divide
AI output now exceeds anyone's ability to understand
Inside large organizations, codebases, specs, tickets, and product documents are now reportedly written entirely by AI with no engineer or manager fully understanding what was generated. Credential-chasing career paths may be training the next generation for roles that no longer exist by the time they arrive.
- Institutional comprehension may lag institutional output by a growing margin
- Inertia, not skill, may be what is keeping some jobs intact
- Individual agency over one's own capability may be the durable asset