The 80/20 of running agent fleets
David Ondrej published his full agentic engineering setup after 6,775 sessions and 2,000+ hours coding with AI. The stack: one open-source interface over every subscription, a self-hosted VPS fleet instead of locked-in cloud platforms, repeated prompts turned into skills, and hard rules about models, reviews and who owns the decisions. Here it is, mapped.
One interface to rule the subscriptions
Main driver is bb, his open-source “agent IDE”: any model, any subscription — Codex, Claude, Pi, Cursor CLI, Grok — in a single UI, with project sidebar, subagent launching and worktrees that remember per-repo settings. cmux (terminal panes + built-in browser) is the second tool, good for small setups and for letting an agent type into the pane next to it, but it doesn’t scale to many projects. Under it all: Ghostty terminal running Herdr — “tmux for agents”, persistent sessions with per-agent state (running / idle / blocked / done). His bet: as manager-and-worker fleets arrive, agent state tracking becomes the core skill within 3–6 months. He also demoed Coral, his own not-yet-open-sourced experiment that queues finished agents by priority instead of random order.
The subscription ladder
Never pay API pricing — “there is no deal there”. Subscriptions are subsidized; stack them by budget. Codex subsidizes harder than Claude (“OpenAI has more compute”). Cursor is his bet to become great value in months, thanks to SpaceX-scale GPUs post-acquisition.
| $10/mo | OpenCode Go — Kimi K3, Grok 4.6, GLM 5.3, DeepSeek V4 Pro. Insane value, small limits, no frontier models. |
|---|---|
| $20/mo | ChatGPT/Codex or Claude Code (Cursor if you want model variety — less subsidized). |
| $30/mo | OpenCode Go + Codex. |
| $50/mo | + Claude Code. |
| $70/mo | + Cursor. Lowest full ladder. |
| $110–210/mo | The big $100–200 Codex + Claude plans — by far the best $/token. Add Cursor only if you can afford all. |
The move of the video: your own agent cloud
Cloud agents are inevitable — but hosted platforms mean hours of environment setup, your secrets and context on their servers, and brutal switching costs. His counter: a cheap VPS running Herdr, reached over SSH from anything — laptop, phone. In the video an agent sets the whole thing up in plain English: installs Herdr, then installs Pi on the box, finds an OpenRouter key on his MacBook by itself, and leaves persistent sessions running that survive a dead laptop. All for a few dollars a month, no lock-in. (The VPS is Hostinger and the video is sponsored — but the pattern is provider-agnostic.)
“You need to fight to set up the first agent. After that, everything else is so much easier.” David Ondrej, 37:40-ish of an hour worth of exactly this
Harnesses, and when to use which
- pi (pi.dev)“The GOAT.” Minimal (4 tools), always YOLO mode, any model or provider. OpenClaw is built on it. Non-negotiable, first install on any new machine.
- Cursor CLIUnderrated: any model, tag skills, and — unlike Claude Code — preset messages you can queue before the agent finishes.
- Hermes / PrimeSelf-learning harnesses for high-uncertainty work: they create skills and improve with you. Test both, pick one. Don’t use them for routine dev — they’ll spray random skills.
- Claude / CodexThe classics for everything else — but alias them (cc, cx) so you never type the YOLO flags daily. “That’s half of agentic engineering: how can you work faster?”
Skills that earn their keep
All free in davidondrej/skills — grab the ones you need, don’t bloat the agent:
| total-review | Launches GPT review + Fable review subagents on your changes, merges both into one deduped list. For medium+ changes, always when cross-model. |
|---|---|
| ask-then-build | Before building, forces the key design decisions on you, with options. “Models are great at implementation. They don’t have taste.” |
| DeepAPI | Deep research, scraping, finding contacts — goes where built-in web search gets blocked. Used daily, whole team on it. |
| global-agent-guardrails | Pre-tool-call hook blocking recursive deletes, git history overrides, password-manager access… “Never run YOLO mode without this.” |
| pro-push | Kernel-level flock: one agent at a time runs merge → verify → push → CI → deploy → health-check. For 15+ parallel agents, otherwise CI clogs. |
The rules, rapid-fire
Do
- Match model to task: Fable to plan, GPT-5.6 Sol max effort for deep bugs, Grok 4.6 for chat (same brains, 2× cheaper), Kimi K3 for frontend.
- When a major model drops, block a full day to use only that model.
- Preset the next 1–6 prompts you already know are coming.
- Document decisions as ADRs in /doc/adr — future agents read why, not just what.
- Give agents a read-only Postgres role on prod: usage-driven development.
- Turn any prompt you repeat into a snippet; dictate with voice (3–4× faster).
Don’t
- Don’t review recursively — one review, fix the worst, ship. Reviews of reviews invent imaginary bugs.
- Don’t let the harness pick subagent models — you’ll get a trash model implementing changes on a codebase you value.
- Don’t use worktrees on small projects; they’re mandatory at 15–30+ parallel agents on big ones.
- Don’t forget to say “don’t add tests” — GPT-5.6 Sol will add hundreds to the smallest repo.
- Don’t give agents write access to prod. One bad call, irreversible.
- Don’t pay API pricing. Ever.
Lectura
The stack is disposable; the posture isn’t
He says it himself: in a month this setup will be different. The durable part is the posture — the person running 6,775 sessions trusts platforms the least: he self-hosts the fleet on a cheap VPS, keeps every architectural decision human-owned via ask-then-build and ADRs, treats model subsidies like airline pricing, and puts a guardrail hook between every agent and every dangerous command. Manager and worker agents are coming, he thinks — but the humans who design when they spawn, and with which model, keep the leverage.