velab.org — Editorial
The umbrella venture, and the signals are describing its own stack back to it.
What the signals say. velab.org's 14 signals are less about a single market than about the portfolio's own operating stack becoming cheap, agent-native, and documented. The single most on-the-nose signal is ibocodes running an entire 2026 product stack on Cloudflare for $5/month — Workers, D1, KV, R2, Queues, Hyperdrive. That is the VE Lab default stack. Apple shipped a native container runtime; Hermes Desktop has a full public tutorial; and the most-repeated business signal is a team of 170 AI agents making every company decision. The umbrella is being handed the playbook for running itself with near-zero human ops overhead.
Directly applicable tools
- Cloudflare full stack at $5/mo — adopt this explicitly as the umbrella infra standard so every venture stops re-deciding it.
- Apple
container — native Linux containers on macOS for local dev without Docker overhead.
- MinerU + Baidu Unlimited-OCR — document AI for any venture that touches unstructured input.
- singleserver (dvassallo) — a reference for giving vibecoded apps real backends cheaply.
- NVIDIA Metropolis VSS 3 + LocateAnything — vision stack for the chipmonk/robotics adjacent ventures.
Competitors & adjacent products
The honest read: velab.org is not yet the umbrella the portfolio assumes it is (per the prior Stack engagement). These signals don't fix that — they make the gap more obvious, because the $5/mo Cloudflare stack and the 170-agent company are exactly the kind of thing a real umbrella would publish as its standard and none of the ventures would have to rediscover.
Recommendations
- Publish the VE Lab standard stack as a single living doc on velab.org — Cloudflare Workers/D1/KV/R2/Queues, Apple containers for dev, Hermes Desktop for agent sessions, MinerU/OKF for ingestion. Stop letting 37 zones re-derive it independently.
- Pilot the 170-agent ops model on one internal function first (support triage or signals ingestion) before claiming it portfolio-wide. The signal is seductive; the failure mode is a fleet of agents with no evals — apply the LLM-as-judge discipline from the freeintelligence feed before scaling.