LLM Wiki — Agentic AI Landscape#
A persistent, compounding knowledge base about the agentic AI ecosystem, built and maintained by an LLM following the llm-wiki-pattern proposed by andrej-karpathy.
What This Wiki Is#
Instead of re-deriving knowledge from scratch on every question (like RAG), this wiki incrementally compiles and maintains a structured, interlinked collection of markdown files. Every source ingested updates entity pages, concept pages, cross-references, and synthesis — so the knowledge compounds over time.
The human curates sources, directs analysis, and asks questions. The LLM does everything else — summarizing, cross-referencing, filing, and bookkeeping.
What’s Inside#
17 sources across tools, standards, methodologies, evaluation, and practitioner insights:
Tools: scion (GCP), kiro (AWS), claude-code (Anthropic), fabric (Miessler), pai (Miessler), paperclip (company-level orchestration), promptfoo (eval tooling), notebooklm (Google Labs)
Methodologies: spec-kit (GitHub, spec-driven development), bmad-method (agile AI-driven development), ten-pillars-agentic-skill-design (Forster)
Standards: agent-skills-standard (agentskills.io), mcp-protocol (Model Context Protocol)
Evaluation: anthropic-eval-guide, evaluating-agent-skills-caparas, promptfoo — from methodology to tooling
Practitioner Insights: ai-technique-podcast, skills-pipeline-sleestk — real-world patterns and skill pipelines
The Emerging Stack#
Five distinct layers have emerged across the 16 sources:
| Layer | Tools | Focus |
|---|---|---|
| Company | paperclip | Org charts, budgets, governance, goal alignment |
| Methodology | spec-kit, bmad-method | Specs, plans, tasks, quality gates, agile workflows |
| Infrastructure | scion | Containers, runtimes, harnesses, isolation |
| Tool | claude-code, kiro | Agentic loop, skills, hooks, MCP, permissions |
| Pattern | fabric, agent-skills-standard | Curated prompts, composable strategies, reusable skills |
Key Themes#
Across 16 sources, eight themes keep surfacing (see cross-source-themes for the full analysis):
- Context beats clever prompting (strongest consensus) — Progressive disclosure, context documents, selective loading.
- Composition over monoliths — Every tool chose small, focused, composable units.
- The human stays in the loop — but how much? — A spectrum from “always interactive” (scion) to “expert collaborator” (bmad-method) to “days of autonomy” (kiro) to “self-modifying” (pai).
- Five orchestration layers emerging — Company, Methodology, Infrastructure, Tool, Pattern.
- Two methodology philosophies — “Specs before code” (spec-kit) vs. “Expert collaboration over autopilot” (bmad-method). Prescriptive vs. scale-adaptive.
- Memory is the unsolved frontier — Persistent context compounds value and errors.
- Open standards winning — mcp-protocol + agent-skills-standard as two-layer open substrate.
- Evaluation is the weakest link — See how-to-eval-a-skill for a practical framework.
Analyses#
- key-insights-agentic-landscape — 10 key insights across the landscape
- cross-source-themes — 8 common themes with evidence tables from all sources
- ten-pillars-evidence-map — How the wiki validates (and challenges) the Ten Pillars framework
- how-to-eval-a-skill — Practical guide: 5 surfaces, 3 tiers, pass@k, CI/CD integration
How It Works#
Three operations:
- Ingest: Drop a source → LLM processes it → creates/updates wiki pages → updates index and log
- Query: Ask a question → LLM reads index, synthesizes answer → optionally files back as analysis
- Lint: Health-check for contradictions, orphan pages, stale claims, missing cross-references
Browse#
- Sources — 17 raw sources that feed this wiki
- Concepts — 18 concept pages covering patterns, standards, and architectural ideas
- Entities — 16 pages for tools, people, and organizations
- Analyses — 4 synthesized analyses filed back into the wiki