This is a publishable agent-orchestration setup, and it's coherent enough to act on. The core claim: eleven agent harnesses (Codex, Claude Code, Grok, Gemini, etc.) each run in their own container under nemesis8, which handles persistent sessions, per-agent identity, telemetry (file changes, container CPU, bandwidth, tokens), and cross-agent session search. Hyperia is the terminal front end — each agent is a visible pane with an MCP server that lets agents drive panes, tabs, and a real built-in browser pane, no extension relay. The pipeline is crawl → embed → search → remember → publish: grubcrawler fetches and cleans to markdown, shivvr embeds, lume does hybrid search with a knowledge graph, ferricula handles long-term memory with decay and consolidation, and nuts.services provides identity plus a public write ledger. Notably, he keeps merges, tags, releases, and installer signing for himself — human-in-the-loop at exactly the irreversible points, which is the part most setups get wrong. The chain of custody question the page raises and doesn't answer: a ledger records that a token wrote something, not which brain held the pen. Eleven containers, same workspace, different brain — and per his own auth history, agents rewrote the security layer. Who reviewed the reviewers?