Agents / agentic engineering198 — 31%
agent design — orchestration — skills — harnesses — multi-agent — coding agents
Evals / evaluation56 — 9%
eval frameworks — LLM-as-judge — agent evals — observability
Enterprise / production46 — 7%
regulated deployments — platform engineering — scaling
Product / startup40 — 6%
shipping AI products — devtool pivots — pricing — GTM
Career / pay / teams37 — 6%
comp models — hiring — which jobs survive — the 10x/1000x engineer
Model routing / LLM ops33 — 5%
routing — serving — latency/cost — gateways
RAG / retrieval28 — 4%
RAG in production — GraphRAG — vector DBs
Realtime / voice20 — 3%
voice agents — interruptions — multimodal
MCP / tool protocols30 — 5%
MCP servers — remote MCP — agent SDKs
Local / edge models12 — 2%
llamafile — Ollama — edge models
Coding agents / dev tools12 — 2%
Claude Code — Codex — Cursor — Devin internals
Security / trust~40 secondary
cross-cutting — permissions — provenance — supply chain
The single largest knowledge batch the fleet has taken in — 43MB of agentic-engineering thinking from the builders themselves (OpenAI, GitHub, Databricks, Resolve AI, ZS, Trunk Tools), synthesised, audited, and wired into the fleet's own review practices. The bill was real; the corpus is permanent; the lessons are already running.
The 114 walled videos + the 259 no-subs are recorded in the manifest — nothing silently dropped. The synthesis read 81 at depth; the full corpus awaits the notebook treatment when you're ready.