06-reference

12 Factor Companies — Jeff Huber

2026-07-21·reference·source: X long-form article·by Jeff Huber (@jeffreyhuber, CEO of Chroma)
ai-era-orgsagentsorg-designharness-engineeringpositioning

Why in vault: Founder shared 2026-07-21. A checklist manifesto for AI-era org design, in the lineage of 12-factor apps → dexhorthy's 12-factor agents. Load-bearing because it is external, quotable vocabulary for what RDCO/CAF already does — validation + sales language, not new instruction.

Bias flag: Huber runs Chroma (vector DB / context infrastructure). "Own your context," "don't let the context vampires in," and "rent your intelligence" are his product's worldview. The advice is still sound — but the emphasis is not disinterested.

The 12 (+1) factors, compressed

Core frame: "The firm of the past was designed around the bottleneck of human execution. The firm of the future is designed around the bottleneck of human taste and judgement… The old firm bundled labor and coordination. The new firm bundles context and authority."

  1. Own your context — full, unlimited access to your data + derived data; write down tacit knowledge so the org is legible to humans AND agents.
  2. Rent your intelligence — keep models swappable; intelligence trends commodity.
  3. Build your tools — the build/buy line has moved; your tooling defends your context.
  4. Buy infra — primitives, not differentiators.
  5. Hire agents before people — new task → try teaching an agent first.
  6. Small, extremely high-taste teams — optimize for the new bottleneck (taste).
  7. Maximize internal information openness — silos are catastrophic for agents; agents are only as good as their context.
  8. Design a compounding learning machine — winning = learning faster; minimize cycle time.
  9. Develop rubrics — instructions + scorecard for critical tasks; iterate the rubric itself.
  10. Reactive + Proactive — background agents with enough org context to propose, not just execute.
  11. Store and learn from production traces — real execution improves agents; near-"recursive self-improvement" for applied builders.
  12. Encode your expertise — agents identify recurring issue patterns, interview you, encode the fix.
  13. (bonus) Think before you AI — your taste is the durable advantage.

RDCO mapping (the reason to keep this)

RDCO/brigade-house already runs ~11 of 12: vault+cellar = #1/#7; Claude-on-subscription = #2; skills/plugins/brigades = #3; Cloudflare/1P/Notion = #4; Ray = #5/#10 (this note was filed by a background agent); eval gates + per-element bars = #9; cellar evals/exemplars/tasting/freshness = #8/#11; competency fill + /improve + /skillify = #12. The gap worth noticing: #11 production-trace learning is our weakest live loop (journals exist; systematic learning from them is manual).

Use for: phData/CAF positioning vocabulary ("rubrics" = eval gates, "encode expertise" = competency cellar, "compounding learning machine" = fill→exemplar→eval→tasting→freshness loop) · Sanity Check evidence (not topic) · the L5 north-star framing.

Cross-links

Why this is in the vault

Mapping against Ray Data Co