The 7 layers of a Claude OS — Karpathy proved CLAUDE.md is the product. Brainfile is what comes next.

Layer 1 — Infrastructure agents 11 agents

Plumbing. Not customer-facing. Examples in the Brainfile spec:

• data_freshness_sentinel — checks every output file against an expected freshness window. If a feed is stale, it raises and re-queues the upstream job.
• deploy_gate — hard gate that runs visual QA before any push to production. Blocks the deploy if the customer view fails any check.
• visual_qa — verifies what the customer actually sees: empty tables, stale timestamps, broken images, mobile overflow, missing CTAs.
• backup_html — versions every dashboard write so a bad generator run is recoverable.
• pii_audit — scans every customer-facing file for addresses, phones, family names, account numbers. Blocks publish on any hit.
• drift_detector — diffs current state against the canonical spec and re-queues any drift.
• backlog_audit — surfaces aging requests so nothing gets forgotten.
• cron_supervisor, secrets_rotator, git_committer, cost_tracker.

Layer 2 — C-Suite agents 5 agents

Strategic. These think about the business, not the code.

• cpo (Chief Product Officer) — owns the roadmap. Reviews every queue item against product strategy.
• cto — owns architecture. Reviews every infra change for security, scalability, drift.
• cmo — owns growth. Drafts content, schedules posts, analyzes funnel.
• vp_customer_success — drafts replies, tracks NPS, surfaces churn signals.
• vp_revenue — owns the revenue gate. Every session must produce at least one revenue-advancing action until ARR exits the danger zone.

These run on a weekly cadence. They write strategic briefs. They populate the priority queue with what the founder is not asking for but should be.

Layer 3 — Algo specialists 4 agents

For the businesses that need them. In the trading stack:

• signal_engineer — proposes new signals from FRED, GDELT, options flow.
• adj_calibrator — backtests adjustment multipliers against 5 years of outcomes, enforces stacking ceilings, never lets a BUY suppressor compound below x0.15.
• shadow_runner — runs new logic in parallel paper-trading for 10+ days before any production wire-up.
• outcome_tracker — closes trades, attributes P&L, surfaces drift in win-rate by regime.

Layer 4 — Skills 30+ skills

Karpathy's repo is mostly skills — small, composable, single-purpose. Brainfile ships skills for:

• Writing — newsletter draft, X thread, landing page copy, refund email, churn-saver.
• Code — generator scaffold, CRUD endpoint, dashboard tab, cron entry.
• Research — competitor scan, pricing audit, SEO keyword pull.
• Ops — Stripe webhook handler, Resend dispatch, domain DNS check.

Skills are the most reusable layer. The same landing_page_copy skill ships across all six businesses; only the inputs change.

Layer 5 — Routines 15+ routines

Anthropic launched Claude Routines in April 2026 — scheduled cloud-side automations. Brainfile ships pre-built routines:

• daily_pipeline — fetch data, regenerate signals, rebuild dashboards, dispatch newsletter.
• weekly_audit — conviction calibration, win-rate review, ADJ effectiveness.
• nightly_visual_qa — runs every site through visual QA, queues fixes.
• monthly_cost_tracker — tallies API spend, flags overruns.
• quarterly_security_audit — rotates secrets, audits permissions, scans for exposed keys.

Routines are the layer that makes the system run while you sleep. Without them, every workflow needs a human to start it.

Layer 6 — Plugins 4+ plugins

External integrations packaged as Claude Code plugins so any agent can call them:

• stripe_plugin — read MRR, churn, customer count, draft refunds, issue partial credits.
• resend_plugin — send transactional + marketing email with per-business sender identity.
• playwright_plugin — browser automation for any task without an API.
• gh_plugin — GitHub PRs, issues, releases, deploy gates.

Layer 7 — Hooks 5+ hooks

Hooks fire on Claude Code lifecycle events. They're how the OS enforces things Claude can't be trusted to remember.

• UserPromptSubmit hook — logs every message to conversation history, runs PII pre-scan.
• Stop hook — auto-commits changed files at session end with a structured message.
• PreToolUse hook on Write — calls backup_html before any HTML overwrite.
• PreToolUse hook on Bash — blocks destructive git commands without explicit approval.
• SessionStart hook — runs the spec drift detector and populates the queue.

Hooks are the quiet layer. Done right, you forget they exist. They are what stops the system from drifting between sessions.