Wiki Log#

[2026-04-07] create | Wiki initialized#

Wiki structure created: CLAUDE.md schema, wiki/index.md, wiki/log.md, directory scaffolding. Ready for first ingest.

[2026-04-07] ingest | Scion Documentation#

Source: https://googlecloudplatform.github.io/scion — multi-agent orchestration testbed by Google Cloud Platform. Created 13 wiki pages:

[2026-04-07] ingest | Kiro Autonomous Agent#

Source: https://kiro.dev/autonomous-agent/ — AWS’s frontier agent for autonomous development tasks. Created 6 wiki pages, updated 1:

[2026-04-07] ingest | Claude Code Documentation#

Source: https://code.claude.com/docs/en — Anthropic’s agentic coding tool. Created 4 wiki pages, updated 2:

[2026-04-07] update | Claude Code Documentation (deep ingest)#

Discovered raw .md endpoints on the docs site. Re-ingested overview, how-claude-code-works, memory, sub-agents, permission-modes, hooks, and skills pages with full content. Updated:

  • claude-code-docs — comprehensive takeaways from actual doc content
  • claude-code — full architecture (agentic loop, tools, memory, permissions, extensibility)
  • mcp-protocol — richer detail on Claude Code’s MCP usage

[2026-04-07] ingest | Anthropic Skills Repository & Agent Skills Spec#

Source: https://github.com/anthropics/skills + https://agentskills.io/specification — open standard for agent skills and Anthropic’s reference implementation. Created 2 wiki pages, updated 3:

[2026-04-08] ingest | The Ten Pillars of Agentic Skill Design#

Source: ~/Documents/2-Areas/agentic-skills-research/agentic-skills-best-practices.pdf (.md version) — research paper by Ian Forster proposing a ten-pillar framework for skill design. Created 3 wiki pages, updated 2:

[2026-04-08] ingest | LLM Wiki (Karpathy)#

Source: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f — the foundational idea file for the LLM Wiki pattern. This wiki is a running instance of this pattern. Created 3 wiki pages:

[2026-04-08] ingest | Fabric#

Source: https://github.com/danielmiessler/fabric — open-source framework with 251+ curated AI prompt patterns and composable strategies. Created 3 wiki pages, updated 2:

[2026-04-08] ingest | Personal AI Infrastructure (PAI)#

Source: https://github.com/danielmiessler/Personal_AI_Infrastructure — personalized AI platform built on Claude Code with memory, skills, goals, and continuous learning. Created 2 wiki pages, updated 1:

[2026-04-08] query | Key Insights: The Agentic AI Landscape#

Synthesized 10 key insights from all 8 sources in the wiki. Filed as analysis: key-insights-agentic-landscape. Covers: three architectural philosophies, the autonomy-interaction spectrum, emerging open standards, progressive disclosure consensus, memory as unsolved frontier, git as universal coordination, skill hierarchy crystallization, security trust models, the personal AI vision, and what’s missing.

[2026-04-08] ingest | How to Evaluate AI Agent Skills Without Relying on Vibes#

Source: Article by JP Caparas — practical guide to skill evaluation with three-tier framework and economics. Created 2 wiki pages, updated 1:

[2026-04-09] ingest | strAIght talk: AI Tips for Amazonians#

Source: Podcast notes from raw/human/ai-technique-podcast.md — Amazon employees sharing AI workflow tactics. Created 1 wiki page:

[2026-04-09] update | Overview#

Updated overview.md with current stats (10 sources, 18 concepts, 10 entities), added “Key Themes” section synthesizing cross-cutting patterns, added ai-technique-podcast to practitioner insights.

[2026-04-09] ingest | Skills Pipeline (Sleestk)#

Source: https://github.com/Sleestk/Skills-Pipeline — chained skill pipelines for YouTube production, SaaS development, and Obsidian. Created 1 wiki page:

  • Source: skills-pipeline-sleestk
  • Key insight: Skills as pipelines — each stage’s output is the next stage’s input. Concrete implementation of progressive disclosure and context management recipes. Ships with inline test prompts.

[2026-04-09] query | Ten Pillars Evidence Map#

Synthesized evidence from all 11 wiki sources supporting each pillar of the Ten Pillars framework. Filed as analysis: ten-pillars-evidence-map. Strongest pillars: Scope (SRP), Modularity, Context Management. Weakest: Versioning, Testing. Identified 4 gaps for v3.

[2026-04-09] query | Cross-Source Theme Analysis#

Identified 8 common themes across all 11 sources with evidence tables and strength rankings. Filed as analysis: cross-source-themes. Strongest: “Context is king” (9/11), “Composition over monoliths” (8/11). Weakest: “Evaluation is weakest link” (4/11 — but universally acknowledged as a gap).

[2026-04-09] query | How to Eval a Skill#

Practical guide mapping Anthropic’s prompt eval methodology onto skills. Filed as analysis: how-to-eval-a-skill. Covers: 5 testable surfaces (routing, tool selection, process, side effects, output quality), 3 tiers, pass@k, eval.yaml format, Claude Code hook integration, CI/CD pipeline.

[2026-04-09] ingest | Anthropic Eval Guide + Promptfoo#

Two eval resources ingested. Created 3 wiki pages:

  • Source: anthropic-eval-guide — canonical eval methodology (SMART criteria, eval types, “volume over quality”)
  • Source: promptfoo — open-source eval CLI (YAML test cases, CI/CD, red teaming)
  • Entity: promptfoo — the closest existing tool to a turnkey skill eval pipeline

[2026-04-09] ingest | Paperclip#

Source: https://github.com/paperclipai/paperclip — open-source orchestration for zero-human companies. Created 2 wiki pages, updated 1:

[2026-04-10] ingest | Spec Kit#

Source: https://github.com/github/spec-kit — GitHub’s spec-driven development toolkit. Created 2 wiki pages:

  • Source: spec-kit
  • Entity: spec-kit
  • Key insight: SDD is progressive disclosure applied to development — each step produces a focused artifact that feeds the next. Agent-agnostic (30+ agents), with 50+ community extensions. Operates at the methodology layer between organizational orchestration (Paperclip) and tool-level execution (Claude Code).

[2026-04-10] ingest | BMad Method#

Source: https://github.com/bmad-code-org/BMAD-METHOD — AI-driven agile framework with scale-adaptive intelligence. Created 2 wiki pages:

  • Source: bmad-method
  • Entity: bmad-method
  • Key insight: “Expert collaboration over autopilot” — agents guide your thinking rather than replacing it. Scale-adaptive planning adjusts depth to project complexity. Party Mode enables multi-persona dialogue in a single session.

[2026-04-11] ingest | NotebookLM Notes Guide#

Source: https://medium.com/@stevenbjohnson/getting-the-most-out-of-notes-in-notebooklm-d9d70316b780 — how-to guide by Steven Johnson on NotebookLM’s notes system. Created 3 wiki pages, updated 2:

  • Source: notebooklm-notes-guide
  • Entities: notebooklm, steven-johnson
  • Updated: context-management — added provenance-as-context-metadata section, NotebookLM’s 5K word limit as concrete context constraint
  • Updated: llm-wiki-pattern — added contrast section comparing NotebookLM’s session-oriented approach to the cumulative wiki pattern
  • Key insight: NotebookLM validates the “human curates, AI processes” principle from a different angle — interactive exploration with provenance tracking vs. structured ingest with automatic cross-references.

[2026-04-14] ingest | Batch 1: Agent Memory & Persistence (4 sources)#

Ingested four sources addressing the wiki’s “memory is the unsolved frontier” gap:

  • mem0-memory-management — Mem0 engineering team: four memory layers, two-phase extraction pipeline (ADD/UPDATE/DELETE/NOOP), LOCOMO benchmarks (93% token reduction), graph-enhanced retrieval
  • continuum-memory-architectures — CMA paper (Logan, arXiv): six formal requirements for agent memory (persistence, selective retention, retrieval-driven mutation, associative routing, temporal continuity, consolidation). CMA won 82/92 trials vs RAG.
  • agent-memory-systems-2026 — Bswen practitioner comparison: four patterns (vector-only, graph+vector, file+DB, hierarchical) with cost estimates and decision framework
  • efficient-memory-architectures — Towards AI guide: four failure modes of flat vector storage, H-MEM, MemGPT (90% token savings), GraphRAG, selective forgetting (RIF formula)

New pages created:

  • wiki/sources/mem0-memory-management.md
  • wiki/sources/continuum-memory-architectures.md
  • wiki/sources/agent-memory-systems-2026.md
  • wiki/sources/efficient-memory-architectures.md
  • wiki/concepts/agent-memory-persistence.md
  • wiki/entities/mem0.md

Updated pages:

  • wiki/concepts/context-management.md — added memory architecture cross-references, linked to agent-memory-persistence
  • wiki/index.md — added 4 sources, 1 entity, 1 concept

[2026-04-14] ingest | Batch 2: Evaluation Benchmarks (4 sources)#

Ingested four benchmark sources addressing the wiki’s “evaluation is the weakest link” gap:

  • humaneval-benchmark — OpenAI (Chen et al., 2021): 164 Python problems, pass@k metric, 0% → 96.3% in 3 years. EvalPlus reveals 7-8 point robustness gap.
  • swe-bench — Princeton (Jimenez et al., 2023): 2,294 real GitHub issues, top 74.4% resolved. SWE-bench Verified (500 tasks) is the standard.
  • gaia-benchmark — Meta/HF (Mialon et al., 2023): 466 questions requiring reasoning + tools + multimodality. Humans 92%, AI <50%. Inverse of most benchmarks.
  • agentbench — Tsinghua (Liu et al., ICLR 2024): 8 interactive environments, multi-turn. Commercial » open-source. General AgentBench shows narrow-to-broad degradation.

New pages created:

  • wiki/sources/humaneval-benchmark.md, wiki/sources/swe-bench.md, wiki/sources/gaia-benchmark.md, wiki/sources/agentbench.md
  • wiki/concepts/agent-benchmarks.md

Updated pages:

  • wiki/concepts/skill-evaluation.md — added benchmark cross-references
  • wiki/index.md — added 4 sources, 1 concept

[2026-04-14] ingest | Batch 3: Multi-Agent Frameworks (4 sources)#

Ingested four open-source multi-agent framework sources, filling the gap where only product-level tools (Scion, Kiro, Claude Code) were represented:

  • autogen-multi-agent — Microsoft Research (56.8K stars): conversation-based coordination, transitioning to graph-based Microsoft Agent Framework. Magentic-One generalist team.
  • crewai-multi-agent — João Moura: role+goal+backstory agents, sequential/hierarchical processes, built-in short/long/entity memory.
  • langgraph-agent-orchestration — LangChain: state-machine graphs, checkpointing (pause/resume), human-in-loop at any node. Most production-ready OSS framework.
  • openai-swarm — OpenAI: educational, radically minimal. Two primitives (routines + handoffs). Not for production but pattern is production-ready.

Convergence signal: both AutoGen (MAF) and LangGraph moving to graph-based workflows — graphs becoming consensus architecture.

New pages: 4 source pages Updated: multi-agent-orchestration concept (added open-source frameworks comparison table), index

[2026-04-14] ingest | Batch 4: Gaps 4-7 (cost, governance, non-code, UX)#

Ingested four sources covering the remaining wiki gaps:

  • agent-cost-economics — Token costs ($80-1500/mo by profile), five waste vectors (60-80% waste), optimization playbook, $5T macro investment thesis, ROI scenarios (3.2% base vs 14.6% optimistic)
  • agentic-ai-governance — Five-pillar governance framework, Shadow AI ($412K/yr), NIST AI Agent Standards Initiative (Jan 2026), EU AI Act, OWASP AIVSS, prompt injection, kill switches
  • agentic-ai-non-code-domains — Six industries (finance 40-60% compliance reduction, legal existential disruption, healthcare, manufacturing, telecoms, transport), SaaS disruption, agent maturity levels 0-4
  • agentic-ux-patterns — Six UX patterns (Intent Preview, Autonomy Dial, Explainable Rationale, Confidence Signal, Audit & Undo, Escalation), phased adoption, service recovery paradox

New pages: 4 source pages Updated: index

[2026-04-14] update | Overview rewrite for 33 sources#

Rewrote wiki/overview.md to reflect expanded wiki:

  • 33 sources (up from 17), 22 concepts (up from 18), 17 entities (up from 16)
  • Six-layer stack (added Memory layer)
  • Four new themes: graph convergence, token economics, governance frontier, beyond-code expansion
  • Original themes strengthened with new evidence (memory no longer “unsolved”, eval no longer “weakest link”)

[2026-04-15] create | Analysis: Memory Architecture Comparison#

Created wiki/analyses/memory-architecture-comparison.md — “How AI Agents Remember: Four Patterns for Persistent Memory”

Synthesizes 8 sources into a comprehensive comparison:

  • Four architecture patterns (vector-only, graph+vector, file+DB, hierarchical) with decision matrix
  • CMA’s six formal requirements mapped against each pattern
  • Four failure modes of flat vector storage
  • Cognitive science memory types (working, episodic, semantic, procedural)
  • Forgetting as design requirement (Bjork, Ebbinghaus, RIF scoring)
  • Benchmark landscape (full-context 72.9% vs Mem0g 68.4% at 93% fewer tokens)
  • Tool comparison (Scion 0/6 CMA → PAI 5/6 CMA)
  • Progression path: start vector-only → add graph → add hierarchy → add forgetting
  • Five open questions for future research

First of five planned analyses from Six Thinking Hats session. Blog post candidate: “How AI Agents Remember: Four Patterns for Persistent Memory”

[2026-04-15] create | Analysis: Multi-Agent Framework Guide#

Created wiki/analyses/multi-agent-framework-guide.md — “Choosing a Multi-Agent Framework in 2026”

Synthesizes 8 sources (4 OSS frameworks + 4 product tools) into a comprehensive guide:

  • Two-tier landscape: product-level (Scion, Kiro, Claude Code, Paperclip) + open-source (AutoGen, CrewAI, LangGraph, Swarm)
  • Decision matrix across 8 factors (production readiness, control, state, memory, debugging, etc.)
  • Graph convergence thesis: AutoGen and LangGraph both moving to graphs as consensus architecture
  • Emerging stack: Paperclip → LangGraph/MAF → Claude Code/Kiro → Scion
  • Multi-agent memory problem: shared memory, conflicting memories, cascading permissions, cost multiplication
  • Progression path: Swarm (learn) → CrewAI (prototype) → LangGraph (production) → MAF/Paperclip (enterprise)

Second of five planned analyses. Blog post candidate: “Choosing a Multi-Agent Framework in 2026”

[2026-04-15] create | Analysis: Cost Optimization Guide#

Created wiki/analyses/cost-optimization-guide.md — “Why Your AI Agent Costs 10× More Than It Should”

Synthesizes 6 sources connecting token economics to architectural decisions:

  • Five waste vectors (60-80% of tokens wasted) with specific fixes
  • Architecture-cost connections: memory choice → token cost, multi-agent → cost multiplication, context management → direct savings
  • Three-tier optimization playbook ordered by ROI (caching → routing → memory architecture)
  • Macro picture: $5T capex, 3.2% base ROI, token explosion, three historical analogies
  • Cost-Quality-Speed triangle: best optimizations improve two dimensions simultaneously

Third of five planned analyses. Blog post candidate: “Why Your AI Agent Costs 10× More Than It Should”

[2026-04-15] create | Analysis: Governance & Safety Overview#

Created wiki/analyses/governance-safety-overview.md — “The Shadow AI Problem: Governing Agents You Don’t Know About”

Connects governance gap to wiki’s existing security models. Five pillars mapped to six UX patterns as user-facing governance layer. Regulatory landscape (NIST, EU AI Act, OWASP, Singapore). Implementation roadmap (4 phases). Key finding: Autonomy Dial = Claude Code permission modes at different layers.

Fourth of five planned analyses. Blog post candidate: “The Shadow AI Problem: Governing Agents You Don’t Know About”

[2026-04-15] create | Analysis: Beyond Code Industry Impact#

Created wiki/analyses/beyond-code-industry-impact.md — “Agentic AI Beyond Software: Six Industries Being Transformed”

Tests whether wiki’s eight themes generalize beyond software development. Six industries analyzed (finance, healthcare, legal, manufacturing, telecoms, transport). Verdict: all themes are universal, but stakes are higher outside code. SaaS disruption thesis. Agent maturity levels 0-4.

Fifth of five planned analyses. Blog post candidate: “Agentic AI Beyond Software: Six Industries Being Transformed”

All five planned analyses from Six Thinking Hats session now complete. Wiki has 9 analyses total (up from 4).

[2026-04-15] lint | Wiki health check#

Fixed:

  • Removed duplicate bmad-method entry in index (was listed twice)
  • Corrected overview concept count: 22 → 20

Stale analyses flagged for refresh (future task):

  • cross-source-themes — written against 11 sources, wiki now has 33. File itself recommends refresh at 20+.
  • key-insights-agentic-landscape — written against 16 sources. Doesn’t reflect memory, eval, multi-agent, cost, governance, or industry sources.

Info noted:

  • 9 thin Scion-specific concept pages (grove, harness, hub, template, runtime, runtime-broker, plugin-system, agent-state-model, kiro-powers) — 1 source each, few inbound links
  • Missing entity pages for Microsoft, OpenAI, LangChain, Princeton, Meta — minor, referenced but not deeply analyzed

[2026-04-15] update | Refreshed cross-source-themes and key-insights for 33 sources#

cross-source-themes.md: Refreshed from 11 → 33 sources.

  • Original 8 themes retained and strengthened with new evidence
  • Theme 1 (Context): now quantified (93% token reduction)
  • Theme 3 (Human-in-loop): formalized as 6 UX patterns
  • Theme 5 (Memory): upgraded from “unsolved” to “understood with clear tradeoffs”
  • Theme 7 (Evaluation): upgraded from “weakest link” to “framework exists”
  • 4 new themes added: graph convergence, token economics, governance, beyond-code

key-insights-agentic-landscape.md: Refreshed from 16 → 33 sources.

  • 6 layers (was 5, added Memory layer)
  • Autonomy spectrum formalized with UX patterns and metrics
  • Memory upgraded from “unsolved frontier” to “understood”
  • Evaluation upgraded with 4 benchmarks
  • 4 new insights: token economics, multi-agent philosophies, governance pillars, remaining gaps
  • “What’s Still Missing” section updated for current state

[2026-04-15] query + create | Conversational query test → gap analysis#

Tested wiki with cross-cutting query: “Build a long-running knowledge management agent — what does the wiki recommend?”

Wiki provided strong architectural guidance across memory (Graph+Vector), coordination (LangGraph), cost ($200-500/mo), trust (Autonomy Dial), and non-code considerations.

Exposed 5 gaps filed as wiki/analyses/wiki-gap-analysis-apr-2026.md:

  1. No “getting started” tutorial (what to choose vs how to wire it)
  2. No LangGraph + Mem0 integration source
  3. No team-level knowledge management agent sources
  4. No knowledge quality evaluation metrics
  5. No agent observability tooling sources

[2026-04-15] ingest | Gap-filling batch: 4 sources addressing high-priority gaps#

Researched and ingested 4 sources addressing gaps identified by conversational query testing:

  • memory-lifecycle-drift — TDS (Nweke, Apr 2026): five lifecycle components (decay, contradiction detection, confidence scoring, compression, expiry). Fills gaps #6-7 (memory quality metrics, regression).
  • langgraph-mem0-integration — DigitalOcean (Payong, Mar 2026): step-by-step LangGraph + Mem0 integration with production considerations. Fills gap #2.
  • multi-agent-observability — Future AGI (Mar 2026): OpenTelemetry spans, debugging patterns, key metrics for multi-agent systems. Fills gaps #5, #10.
  • ai-environmental-impact — arXiv (Jegham et al., May 2025): energy/carbon/water benchmarks across 30 models. o3 = 70× GPT-4.1 nano. Claude 3.7 Sonnet most eco-efficient. Fills gaps #12-13.

Wiki now has 37 sources. Gaps addressed: #2, #5, #6, #7, #10, #12, #13.

[2026-04-15] create | Analysis: Getting Started Guide#

Created wiki/analyses/getting-started-guide.md — “Building Your First Long-Running Agent”

Five-phase guide synthesizing wiki’s consensus recommendations:

  1. Single agent + Mem0 memory (Week 1)
  2. Memory lifecycle: decay, contradiction, confidence, compression, expiry (Week 2)
  3. Human-in-the-loop: Plan & Propose → Act with Confirmation (Week 3)
  4. Observability: OpenTelemetry spans + key metrics (Week 4)
  5. Scale to multi-agent (Month 2+)

Includes cost budgets per phase ($80-500/mo), common mistakes, and what’s still not covered. Fills gap #1 from wiki-gap-analysis. Wiki now has 11 analyses.

[2026-04-15] ingest | Multi-agent advanced gaps: 2 sources#

  • crewai-production-guide — Practical crew patterns (content pipeline, customer support, event-driven), production deployment, cost tracking. Fills gaps #9 (CrewAI examples), #11 (role design patterns).
  • shared-agent-memory — Shared memory layer pattern (solution store + write hook + retrieval step + threshold), conflict resolution status (implicit via recency, explicit remains unsolved), partial/failed solution storage. Fills gaps #14, #15, #16.

Wiki now has 39 sources. Multi-agent advanced gaps: 5/5 addressed (though explicit conflict resolution confirmed as open problem across all sources).

[2026-04-15] create | Analysis: Hybrid Memory Architecture (Wiki + Mem0)#

Created wiki/analyses/hybrid-memory-architecture.md — “Combining Wiki + Mem0”

Novel architecture synthesized from 7 wiki sources (not documented in any single source):

  • Layer 1: Mem0 as fast memory (working + session + user) — sub-second retrieval, automatic forgetting
  • Layer 2: Wiki as deep memory (semantic + organizational) — synthesis, provenance, git history
  • Layer 3: Bridge — bidirectional sync (wiki→Mem0 index extraction, Mem0→wiki consolidation)
  • Achieves 5/6 CMA compliance (matching PAI) with full transparency
  • Implementation in 5 phases, $1-55/month cost estimate
  • Directly supports content pipeline goal (interactions → Mem0 → wiki → blog posts)

Wiki now has 12 analyses.

[2026-08-11] ingest | Knowledge Graph Validation: Hybrid Human-LLM Workflows#

Ingested research paper by Tsaneva et al. (Information Processing & Management, 2025) on combining LLMs and human-in-the-loop for knowledge graph validation. Empirical study of nine workflows across four automation levels (human judgment, AI assistance, human verification, full automation) tested on CS-KG (3.6K triples from 6.7M publications).

Key finding: Disagreement strategy (workflows 5 & 6) optimal — human validates only when automated methods conflict. Achieves F1 +5% with <13% human effort, no precision/recall tradeoff.

New pages created:

Updated pages:

  • skill-evaluation — added KG validation as concrete instance of three-tier evaluation framework

Connections to existing wiki:

Wiki now has 40 sources, 24 concepts, 23 entities.

[2026-08-11] ingest + update | AI-Powered Dependency Graph Analysis (Cross-Domain Pattern Identified)#

Ingested CELSO consulting guide on AI-driven code dependency graphs for impact analysis, CI/CD integration, and compliance. Key finding: Same three-tier validation pattern as kg-validation-hybrid-workflows, applied to code systems instead of knowledge graphs.

Breakthrough: Disagreement strategy + three-tier escalation is domain-agnostic pattern, not KG-specific.

New pages created:

Updated pages (to reflect cross-domain pattern):

  • human-in-the-loop — added cross-domain pattern table showing disagreement strategy across KG validation, code impact analysis, and agent orchestration
  • skill-evaluation — renamed section to “Concrete Instances”, added both KG validation AND code dependency analysis as examples of three-tier framework

Pattern unlocked: Graph-based systems benefit from three-tier validation (automated → AI → human on disagreement):

  • Domain 1: Knowledge graphs (fact triples) → F1 +5%
  • Domain 2: Code dependency graphs (artifact relationships) → 40% fewer outages
  • Domain 3: Agent task graphs (via LangGraph) → human-in-loop at uncertain nodes

This strengthens wiki’s architecture: pattern is platform/domain agnostic, applies wherever you have:

  1. Large interconnected graphs
  2. Automated baseline validators
  3. High cost of false negatives
  4. High human annotation cost

Wiki now has 41 sources, 25 concepts, 24 entities.

Backward Vibe Coding Lessons K10s