Research

1.1Memory and context

1. 2026-11-03R002The 5 layers of agent memoryWorking, episodic, semantic, procedural and meta-memory. Each layer fails the same way: the store exists, but the read, dedupe, reflection or garbage-collection step that keeps it useful is missing.
   agent-architectureagent-memorycontext-engineeringself-improving-agentsagent-governance

• AlsoR010 PageIndex: reasoning-based retrieval for long documents: retrieval as navigation: read a document's map, then only the pages it points to; "similarity is not relevance"
• AlsoR006 6 must-know terms for AI agent harnesses: context management and checkpointing as the harness's per-step context and resumable state
• AlsoR005 Second Brain OS: the five-layer map of AI agents: context as a budget; filesystem as long-term memory, window as working memory
• AlsoR004 The best Claude setup is no setup: argues for turning memory off and wiping instructions after model upgrades

1.2Self-improvement and learning loops

• AlsoR002 The 5 layers of agent memory: reflection pass writing procedural memory; commenters argue it needs eval and human gates

1.3Governance of agent state

• AlsoR005 Second Brain OS: the five-layer map of AI agents: records contradictions with dates and sources instead of overwriting; little on authority or audit
• AlsoR002 The 5 layers of agent memory: open gap: recall vs. permission, and audit of offline memory rewrites
• AlsoR004 The best Claude setup is no setup: periodic manual wipe as governance, with no eval gate

1.4Mental models of the agent system

1. 2026-11-04R005Second Brain OS: the five-layer map of AI agentsAn agent is the model plus five layers, each answering one question: context (what it sees), loop (who decides), gate (who sorts), harness (what it can reach), evals (how you know). A failure in one layer looks like another, so name the layer before blaming the model.
   mental-modelsagent-architecturecontext-engineeringharness-configevalslearning-resources

• AlsoR006 6 must-know terms for AI agent harnesses: six harness concepts as a production vocabulary; overlaps R005's loop, gate and harness layers

1.5Context economics and caching

• AlsoR010 PageIndex: reasoning-based retrieval for long documents: tree as cached prefix, fetched pages as the tail; reads a few pages instead of a whole PDF, 2.1× to 16.6× cheaper by vendor figures

2.1Developer infrastructure under agent load

1. 2026-11-03R001Scaling Git at OpenAI (Pando)Agent fan-out broke stock Git hosting. Dropping repo-wide ref snapshots (moving to causal consistency) and dropping per-client packs (reusing cached packs via packfile-uri) cut median push latency about 80% at about 1 PB/day.
   dev-infragitmonorepoagents-at-scaledistributed-systemsformal-verificationai-built-software

2.2Consistency and storage tradeoffs

• AlsoR001 Scaling Git at OpenAI (Pando): causal consistency instead of repo-wide snapshots
• AlsoR002 The 5 layers of agent memory: contradiction handling in semantic memory

3.1Agents building production systems

• AlsoR012 AI crash projects for public good: Mollick's post and the evidence behind it: delegating the work erases the learning; cites Shen and Tamkin 2026, where developers who learned a library with AI scored worse unaided
• AlsoR001 Scaling Git at OpenAI (Pando): see its summary's notes on AI-built software

3.2Formal methods for AI-written code

• AlsoR001 Scaling Git at OpenAI (Pando): see its summary's notes on formal verification

4.1Docs-as-code and project memory

• AlsoR005 Second Brain OS: the five-layer map of AI agents: LLM-wiki second brain: nothing ingested until linked, index.md + log.md, lint skills

4.3Retrieval over long documents

1. 2026-11-06R010PageIndex: reasoning-based retrieval for long documentsReplaces chunk-and-embed RAG with a per-document section tree. An agent reads the tree, then fetches tight page ranges and cites pages. In the repo's own 10-K notebook it scored 100% at about $110 per 1,000 queries; the best vector run was 92% at $134. Wins on long, structured documents; costs several strong-model turns per query and depends on a correct outline.
   retrievallong-documentscontext-engineeringknowledge-managementagent-harness

5.1Harness tips

1. 2026-11-04R004The best Claude setup is no setupConfig written for older models rots, so wipe instructions, skills and memory, default to Auto, prompt with Goal + Context, and ask for HTML pages instead of .docx. Sound principle (prune and re-baseline after model upgrades), weak procedure: no before/after eval, and it deletes project facts and rationale along with obsolete capability patches.
   harness-configpromptingcontext-engineeringagent-memory

• AlsoR006 6 must-know terms for AI agent harnesses: what the harness does, as opposed to how you configure it; permissions decide when Auto is safe
• AlsoR005 Second Brain OS: the five-layer map of AI agents: four rings built outside-in; approval prompts only for irreversible actions; harness bets expire, re-audit quarterly

6.1Harness building blocks

1. 2026-11-04R0066 must-know terms for AI agent harnessesAgent = model + harness; six concepts: agent loop, tool calling, context management, checkpointing, permissions and sandboxing, guardrails. Good vocabulary, shallow treatment. The comments add what's missing: observability and trajectory evals, and the point that checkpoint-and-resume is only safe when write tools are idempotent.
   agent-harnessagent-architectureagent-reliabilityevalscontext-engineering

• AlsoR012 AI crash projects for public good: Mollick's post and the evidence behind it: the model phrases, deterministic code decides; AMIE's live physician gate and what replaces it at scale
• AlsoR010 PageIndex: reasoning-based retrieval for long documents: retrieval as three tools: browse, get structure, get pages; grounding prompt forces page citations
• AlsoR005 Second Brain OS: the five-layer map of AI agents: five-layer map; its evals layer is the piece R006's commenters say is missing

7.1Learning, advice and public-good deployment

1. 2026-11-06R012AI crash projects for public good: Mollick's post and the evidence behind itMollick wants national crash projects in tutoring, health second opinions, financial literacy and government services using current models. The evidence says the bottleneck is the harness (steering toward augmentation, eliciting context, cheap oversight), not the model: students who let AI write kept nothing, and the medical agent was safe only with a physician watching live.
   ai-impacthuman-learningpublic-goodhuman-oversightevalsprompting

• AlsoR014 MIT Technology Review's Intro to AI mini-course: a general-audience AI literacy course; its verify, iterate and know-when-AI-is-the-wrong-tool habits are what R012's harness lesson asks users to bring

8.1AI literacy curricula

1. 2026-11-07R014MIT Technology Review's Intro to AI mini-courseA six-part beginner course: definitions, a hands-on chatbot exercise, how to talk about AI (including with kids), traps (bias, hallucination, unreliable detection, deepfakes), work and regulation. Durable on judgment (iterate and correct, a fit-for-purpose rule, verify, don't anthropomorphize); dated on tools (no agents or reasoning models) and US policy. The summary adds a checklist for an updated curriculum.
   ai-literacyeducationlearning-resourcesai-impactai-policyhuman-oversight