People keep asking how Ran gets through it all — book(s), Automaze (~30 people), VarOps climbing out of stealth, open source, this show. This episode is the answer: setup, then time, then how the companies actually run.
The stack (technical)
Cloud-local hybrid. Debate rages about local vs cloud coding environments. Ran’s middle path: a dedicated home-office Mac that’s his alone, reachable from anywhere — “local” when he’s exploring, remote when he needs to. Full config, room to experiment.
Harnesses by stage.
- Exploratory / figuring-it-out: Factory’s Droid as the go-to harness; Paseo (Cursor-style local IDE that speaks many harnesses, including Droid) for ad hoc work and even mobile long-running sessions.
- Mature product / feature-fix mode: Cloop — Automaze’s internal Devon-style environment (shipping later). Brainstorming module, Ladybug for production bug hunt → fix → PR (human always in the loop at PR).
- Sessions stay consistent: whether work starts from MUXI, Droid, or Cloop, coding funnels into the same Droid harness.
Assistant layer. MUXI is the personal OS (he built it; he can extend it). Claude Desktop is the client, wired over MCP — Claude as brainstorming partner with MUXI underneath. Morning briefings, Slack triage, email drafts, meeting prep: automations on MUXI.
Everyday tools. Apple Reminders + Notes. Superhuman for email (back after a years-long break). VS Code mostly as a word processor / book-writing environment, not a daily IDE. Hosting for his own projects: Cloudflare, Hetzner, DigitalOcean (clients get whatever fits).
Self-host almost everything — not to save $20/month (token spend is huge), but for data ownership and the option to touch the raw DB when needed.
Time: classify, don’t type
The job used to be doing the work. Now it’s deciding which bucket the work goes in — Bezos-flavored: a small number of high-quality decisions, then time left for building and experimenting (including things that never ship).
~20–30 decisions a day, almost all classifications:
- rails vs agents
- human vs automated vs kill the process
Anything done more than a few times gets automated into one of three shapes: fully deterministic (no AI), hybrid (workflow with a fuzzy AI step), or agent-shaped when judgment is the point.
Most hours go to thinking: stripping products to first principles, shrinking codebases to cut maintenance surface, deciding what not to build. Hours (sometimes days) of brainstorming and speccing in Claude before a line of code — often with GPT and Perplexity in the loop for research/pushback while Ran stays the human in the middle. Generation is cheap; judgment isn’t. He reads agent-produced code and comments heavily.
Email: closed by default, opened twice a day around the briefing. Short VIP list for urgent operational paths (e.g. accountant), not a status ranking.
Automations stay gated on purpose
Not because AI “isn’t ready.” Because every client is an edge case. AI clears dirty work; it doesn’t replace him on partner/client communication. The gate is a judgment dial, not a maturity cop-out.
How Automaze and VarOps actually run
Lean founder ≠ solo. Partners Anton and Alex run the team and pipeline. Ran’s focus: customers in/out, PRDs, product shape — then handoff, with a jump back for the last ~10% and forks in the road.
Leads: ~99% inbound — open source, podcast, newsletter, book(s). Not a quarter-long growth hack; fifteen years of giving things away.
FDE by design. Automaze has always embedded in client environments. Ran personally FDEs new clients early — spec, process, lived context — then hands to the team. Not insecurity about company size; you can’t talk confidently (or improve process) about what you haven’t felt. Small number of high-quality decisions requires understanding first.
The point
Machine, lists, pipeline — all scaffolding to protect the context in his head so he can classify all day. That’s the job now.
Newsletter / PDF version teased via aroussi.com.
