homeblogabout

  • rss

  • twitter

  • linkedin

© 2025

Field Notes

Field Notes are fast, from-the-trenches observations. Time-bound and may age poorly. Summarized from my real notes by . Optimized for utility. Not investment or legal advice.

Notebook background
░░░░░░░▄█▄▄▄█▄
▄▀░░░░▄▌─▄─▄─▐▄░░░░▀▄
█▄▄█░░▀▌─▀─▀─▐▀░░█▄▄█
░▐▌░░░░▀▀███▀▀░░░░▐▌
████░▄█████████▄░████
=======================
Field Note Clanker
=======================
⏺ Agent start
│
├── 6 data sources
└── Total 62.3k words
⏺ Spawning 6 Sub-Agents
│
├── GPT-5: Summarize → Web Search Hydrate
├── GPT-5-mini: Score (Originality, Relevance)
└── Return Good Notes
⏺ Field Note Agent
│
├── Sorted to 2 of 7 sections
├── Extracting 5 key signals
└── Posting Approval
⏺ Publishing
┌────────────────────────────────────────┐
│ Warning: Field notes are recursively │
│ summarized by agents. These likely age │
│ poorly. Exercise caution when reading. │
└────────────────────────────────────────┘

Field Notes - Dec 04, '25

Executive Signals

  • Determinism is the new autonomy: freeze agent flows to code; spend AI on fuzz
  • Evals over vibes: prompts as compiled artifacts; releases gated by measured lift
  • Soak before scale: progressive rollout with error budgets and evidence retention
  • BYOK or bust: enterprise trust comes from keys, logs, and server guardrails
  • AEO is the new SEO: machine-readable feeds outrank pixel soup for answer engines

CEO

Center of Excellence, Then Cascade

A small AI CoE de-risks primitives (streaming, routing, evals, safety, RAG) in 15–20 days, then propagates patterns through template repos and playbooks. Pair-led spikes teach the org while keeping standards consistent.

  • Charter a 3–5 person CoE with a batteries‑included template repo
  • Run pair-led hackathons in phase one; cascade ownership in phase two
  • Publish short playbooks and require reuse across teams

Guardrails for Business Commitments

Models should never finalize prices, inventory, or contracts. Gate binding actions behind server policies and human approvals. Scope tools/data per tenant and watch for anomalies to protect margin and trust.

  • Require server-side approvals for binding actions
  • Enforce per-tenant scopes, rate plans, and abuse thresholds
  • Alert on suspicious discounts and unusual token bursts

Reversible Bets, Not One-Way Doors

Frameworks and models shift quarterly. Keep a slim interface around LLM calls, maintain 90‑day swap plans, and avoid deep bets. Prototype fast with a single vendor, then diversify and run production in your cloud.

  • Abstract prompts/orchestration; keep two runner‑up models warm
  • Pilot hosted; operate production in your cloud with canary failover
  • Treat multi-model and BYOK as table stakes at scale

Budget the Last 20% Like It’s 80%

AI automations become expensive in integration, compliance, and edge-case handling. Progress should be measured by error budgets and auditability, not demo flash.

  • Name a single DRI with a responsiveness SLA
  • Gate go-live on error budgets and audit/compliance checks
  • Track demo‑to‑production time and edge‑case burn‑down

Marketing

Build a Generative‑Web Layer Beside Your Site

Agents and LLM crawlers prefer dense, machine-readable facts. Ship an ai‑feed (Markdown/JSON) that mirrors human pages with specs, FAQs, pricing context, and citations. Track LLM citation share as a first-class metric.

  • Add a sitewide ai‑feed and link via header or rel=alternate
  • Log GPT/Perplexity/Gemini hits; monitor citation share weekly
  • Refresh only on material
PreviousDec 3, 2025
NextDec 5, 2025
Back to Blog