The Doctrine — what Nate actually teaches, in 12 claims
Three and a half years, 582 posts, distilled to the load-bearing claims. Each one links to the topic page that carries the receipts. This is the page to reread when you're going in circles: if a project violates one of these, that's usually why.
The stack
1. The harness beats the model. The same model scores 78% or 42% depending on the scaffold around it. Loop structure, memory strategy, tool registry, constraints, verification — that's where outcomes are decided. Stop shopping for models; audit the harness. → Agent Architecture, Coding-Agent Practice
2. Memory, not intelligence, is the binding constraint. Intelligence scaled ~60,000x since ChatGPT; memory infrastructure maybe 100x. Every fresh chat that re-derives what you already knew is the tax. The fix is architectural: owned, queryable state outside any context window. → Memory & Second Brain
3. Context windows are a lie. A 200K-token window behaves like ~20K of reliable attention; more documents can make answers worse. Compute context fresh each call from structured state — don't pin everything into a mega-window. → Memory & Second Brain
4. Open Brain, not Second Brain. Nate explicitly retired the human-facing Notion/Obsidian pattern in March 2026: knowledge stores must be built for agents to query (Postgres + pgvector over MCP, any client reads/writes the same store), with humans served derived views. Your notes app is not your brain; your agents' database is. → Memory & Second Brain
5. Prompting graduated into context engineering. The literal prompt can be a fraction of a percent of what the model processes. What matters: the deterministic context you author (role, contract, constraints, format-last), and the probabilistic context the agent discovers. Order matters mechanically — early tokens steer, format specs act as stop-tokens, guardrails need re-pinning as chats age. → Prompting
6. Constraints beat instructions. Spend more words on what to exclude than include; bake self-scoring rubrics into the prompt so the model grades its own draft before a human sees it. Human review does not scale to AI volume — build the reviewer in. → Prompt Packs, Reliability
7. Hallucination is architecture, not a bug. Default behavior of next-token prediction; no model release removes it. Verification must be a first-class design feature: grounded retrieval for anything high-stakes, second-model cross-checks, track hallucination like uptime. Green tests that mask real defects are the same failure. → Reliability
8. Single-agent-first; multi-agent only for genuine parallel expertise. Multi-agent costs 3–10x in tokens and time. Dumb agents + smart orchestration beats clever autonomous agents. Route by task shape: iterative/collaborative work to Claude-style agents, delegate-and-walk-away parallel work to Codex-style sandboxes. → Agent Architecture
9. MCP is for background intelligence, never hot paths. Hard numbers: 300–800ms latency, token blowups up to 236x, real cross-tenant breach from token passthrough. Protocol choice is an architecture decision with a security budget, not a checkbox. → Agent Architecture
10. The bottleneck moved from execution to judgment. Execution is becoming free; clarity, taste, specification, and distribution are what's scarce. Define what "correct" means — which claims the system may make, what evidence backs each, whether a confident wrong answer or a refusal is worse — before building anything. → Operator Playbooks
11. Your data pipe, not the model, is your moat. The Narrow-Pipe Law: a noise-free, purpose-built input stream turns a commodity model into a bespoke analyst. Chat is a demo; value lives in code execution, cross-tool orchestration, and persistent reusable threads. → Operator Playbooks
12. Reading code (and work) overtook writing it. AI floods repos and workstreams with output nobody vets. The durable human skill is tracing, auditing, and owning invariants — engineering rigor goes up, not away. → Coding-Agent Practice
What this means for the Ketron fleet
The uncomfortable convergence: this corpus independently validates the direction already in the canon — harness audits, judge hooks, owned memory (KWM/Open Brain), verification before DONE — and indicts the same failure modes the canon logs (re-derivation, green-test theater, artifact sprawl instead of one queryable store). Claim 4 is the sharpest edge: the wiki you are reading is a human-facing view. The agent-facing store (the corpus repo + catalog it's built from) is the part that compounds. Point Rain, NEXUS, and every future session at the store, not the view.