Inside the Nerve Center
Tier 3

Inside the Nerve Center

How my production AI organization works today: the architecture, the story, orchestration, memory, the cockpit, and how I talk to it.

20 articles113 min total

Start reading →

Overview

  1. The architecture in one pageOperator, cockpit, bridge, Claude Code sessions, memory, Postgres, watchdogs and one send gate: the Nerve Center on one page.5 min read
  2. How the system got builtFrom a chat bot in early 2026 to a production AI organization: breadth, then consolidation and safety, then depth.6 min read

Orchestration

  1. The CEO agentThe orchestrator I talk to: how it routes work to domain agents and specialists, and what it may never do alone.6 min read
  2. Domain agentsOne persistent agent per area of life and work, each with its own folder, state, rules and live session.5 min read
  3. The agent bus: how agents talkFrom a synchronous ask that could silently return nothing to a durable bus that persists, retries and dead-letters loudly.5 min read
  4. Fleets: parallel build agentsMany build agents at once: a worktree supervisor, a queue drainer, off-plan models for fleet work, and the orphan-process lesson.6 min read
  5. Model routing in practiceA budget router in April, one registry per tier, automatic model adoption, an explicit-only top tier, and a local model that crashed the box.8 min read
  6. Expert agents: route by topicA knowledge base plus a routing table, so the main agent calls the expert instead of answering from its own head.5 min read
  7. The executive assistantInbox triage, calendar merge, draft-first replies through the send gate, a pull-based Needs-you view, weekly planning, and an iMessage channel.5 min read

Memory and context

  1. The second brain: logs, rollups, journalsThe system's memory of its own work: per-project logs, a daily rollup, a project registry and a compiled daily journal.6 min read
  2. Memory: graph plus vectors plus keywordsHybrid retrieval over a graph store, a vector store and a keyword index, and why relationships need a graph.6 min read
  3. The CRM: every person becomes a fileMarkdown person files with machine facts kept apart from real human events, built from mail, calendar and meetings.6 min read
  4. Personal context filesFour files that tell every agent who you are, how you decide, how you write and what you never do. Plus five skills.5 min read
  5. Open loops: tracking every promiseA ledger of what the system promised to do, surfaced at every session start and checked by watchers where a machine can check it.7 min read

The cockpit

  1. The cockpitA web app with a live pane for every agent session, an iPhone app around it, and a design system shared with this site.5 min read
  2. What the cockpit rebuilds taught meReplaying a terminal in a browser broke everything. What replaced it, and why syncing intent beats syncing state.6 min read
  3. The bridgeThe service between the cockpit and every agent session, and how one careless restart once killed them all.5 min read

Interfaces and tools

  1. Talking to it from your phoneTelegram as the first control plane, a write-ahead journal so no message is dropped, and an iMessage channel for the assistant.5 min read
  2. Voice in and outSpoken summaries that fire without the model remembering, local text-to-speech, a read-aloud button in the cockpit, and dictation that doesn't invent words.6 min read
  3. The agent browserWhen an agent needs a real browser: Playwright, a signed-in profile, per-site adapters, and the rule that a human finishes the risky part.5 min read