Scout

90 concepts known

Ledger

What the scout will never brief you on again. Everything here is assumed known, so new terms are measured against it.

Prompting & context 10

  • Prompt engineering
  • Context engineering (context management, context windows)
  • Context plumbing (infrastructure for routing live context between sources and agents)
  • System prompts
  • Few-shot / zero-shot prompting
  • Chain-of-thought prompting (CoT)
  • Prompt caching
  • Structured output (JSON mode, constrained decoding)
  • Prompt injection / jailbreaking
  • Cryptographic context injection (encrypted prompt injection)

Agents & orchestration 42

  • AI agents / agentic workflows
  • Agent harnesses
  • Tool use / function calling
  • MCP (Model Context Protocol)
  • MCP Apps (interactive UI extension for MCP)
  • MCP Tasks (Tasks extension, call-now fetch-later pattern for long-running MCP calls)
  • MCP Skills Extension (SEP-2640, Skills over MCP, skill:// resources, skills/list, skills/get)
  • MCP extensions framework (versioned extensions map, reverse-DNS extension IDs)
  • Client ID Metadata Documents (CIMD, stateless OAuth client identity for MCP)
  • Multi Round-Trip Requests (MRTR, MCP's stateless server-request pattern)
  • Enterprise-Managed Authorization (EMA, ID-JAG, Identity Assertion JWT Authorization Grant, IdP-mediated MCP server access)
  • ACP (Agent Client Protocol)
  • A2A (Agent-to-Agent protocol)
  • Model Hardware Standard (MHS, physical-world hardware protocol for agents)
  • Legal Context Protocol (LCP, legal-context.json, consent/dispute-resolution layer for agentic commerce)
  • Multi-agent orchestration / subagents
  • Computer use (GUI agents, browser agents)
  • Agent skills
  • Human-in-the-loop
  • Agent memory (persistent memory, memory files)
  • Harness-managed virtual memory (ClawVM, agent context paging)
  • Recursive Language Models (RLM, recursive self-calling over externalized context via REPL)
  • Agent sprawl
  • Meat proxy (blindly relaying unread/unverified AI output)
  • Loop engineering
  • Eval engineering (evals engineering, versioned/gated eval suites as production infrastructure, successor framing to loop engineering)
  • Graph engineering (execution graphs, agent topology as a graph, successor framing to loop engineering)
  • Managed agents (cloud-hosted agent runtime, server-hosted agent sandbox)
  • Intent engineering (encoding agent objectives/values/trade-offs beyond prompt and context)
  • Thin agents (ontology-based semantic layer, thin-agents-on-shared-substrate pattern)
  • Harness engineering (Agent = Model + Harness)
  • Attention interface (attention-coordination layer)
  • Meta-harness (harness of harnesses, cross-harness orchestration layer)
  • Know Your Agent (KYA, agent identity verification, agent KYC)
  • Agentic engineering (agent engineering, successor framing to vibe coding)
  • Skill engineering (discipline of designing/evaluating agent skills with with-skill/without-skill evals, sibling framing to harness engineering and eval engineering)
  • Advisor pattern (advisor strategy, two-model pattern, executor-advisor pattern, cheap executor escalating to a non-acting advisor model)
  • Agent-scale version control (agent-scale source control, next-gen SCM for AI agents, server-authoritative history/merge + partial checkout + ephemeral agent workspaces)
  • Personal Agent Protocol (PAP, Sierra/Meta open standard for personal agent-to-business identity, consent, and access scoping)
  • WebMCP (Web Model Context, MCP-B synthesis, browser-native tool-exposure standard for in-page AI agents)
  • Claws (Claw, personal agents, general agents — self-hosted always-on personal AI agents reached over messaging, with scheduling and persistent memory; OpenClaw, NanoClaw, ZeroClaw, IronClaw, PicoClaw)
  • Harness² (recursive agent harnessing, two-level task-level/harnessing-level recursion for self-improving a frozen agent's harness without retraining weights)

Retrieval & knowledge 6

  • RAG (retrieval-augmented generation)
  • Embeddings
  • Vector databases / semantic search
  • Reranking
  • Knowledge graphs
  • Ontology-driven guardrails (semantic guardrails, vocabulary ontologies for agents)

Models & training 15

  • Fine-tuning (SFT, LoRA)
  • RLHF / RLAIF
  • Distillation
  • Quantization
  • Mixture of experts (MoE)
  • Reasoning models / test-time compute (inference-time scaling)
  • Speculative decoding
  • Multimodal models
  • Diffusion models
  • Transformers / attention
  • Compressed sparse attention (CSA)
  • Vibe-proving (AI-assisted mathematical proof discovery via generate-referee-repair loop)
  • System One models (decision models, non-autoregressive typed-decision models trained via RLCD)
  • Inoculation prompting (inoculation training, inoculation midtraining, inoculation adapters, selective generalization via inoculation)
  • Opaque recurrence (recurrent depth, looped transformers, latent reasoning loops)

Interpretability 3

  • J-Space (global workspace, J-Lens)
  • Introspection adapters (IA, Diff Interpretation Tuning)
  • Natural language autoencoders (NLA, activation verbalizer/reconstructor)

Evaluation & safety 14

  • Evals / benchmarks
  • LLM-as-judge
  • Guardrails
  • Agent behavioral contracts (ABC, design-by-contract for agents)
  • Red teaming
  • Hallucination
  • Alignment / Constitutional AI
  • Model cards
  • Double-blind AI evaluations (confidential evaluation, cryptographic evaluation boxing)
  • Malicious agent skills (ToxicSkills, AST01, agent skill supply-chain attacks, SKILL.md poisoning)
  • Cross-agent privilege escalation (cross-agent config tampering, confused deputy in multi-agent coding setups)
  • Governance decay (context-compaction erosion of in-context safety/policy constraints, Compaction Cliff)
  • Metagaming (verbalized metagaming, oversight gaming, CoT reasoning about grading/reward/monitoring)
  • Preserved thinking (anti-distillation safeguard, immutable thinking-block context binding)