Dark factory at $0.015/task: models write, scripts decide
Oct 2 Nalivator3000: solo Archon/Paperclip/LiteLLM dark factory — ~$0.015/task, ≤4-file cuts at 40–50% one-shot; silent credit fallback, alert-to-agent trap, 14-day improver cooldown, script-owned merge.
Source: I run a one-person "dark factory" of AI agents for $0.015 per task. Here is what broke.
Nalivator3000’s Oct 2 field note is a month of running a one-person dark factory against LexyBooster (and a side Godot game) without pretending the hard part was the model. Stack is deliberately boring: one Hetzner VPS (~$12.60/mo), Archon for workflows, Paperclip for budgets/schedules, LiteLLM routing free models first (Groq/Cerebras) → Gemini Flash-Lite → Claude subscription for plan/review. Merge authority is a script, never a model: strong-reviewer approve, green checks, no deleted definitions, no forbidden paths.
Receipts after a month: typical one-to-two-file+test tasks cost about $0.015 in API calls (cheapest ~$0.007). The biggest saving was not a smarter model but smaller tasks — epics cut to at most four code files let a single cheap call land a working patch on 40–50% of tasks. The factory shipped six Godot multiplayer epics plus SEO pages with the human mostly out of the write loop.
What broke was ops, not prompt craft — and the fixes were harness changes: (1) prepaid credits ran dry for two days while the pipeline silently fell back to the subscription model and kept merging → gateway-log monitor for “credits depleted”; (2) a PR waited two days because the alert went to an “agent that fixes alerts” that cannot merge config → money/merge alerts skip agents and page the human; (3) the improver re-touched the same module three nights running after seeing its own commits → 14-day cooldown per module; (4) cheap agents spun up to 33 calls on hard tasks → one-shot + two quick repairs, then subscription model; cheap only when the subscription window is closed. Lesson he keeps: let models write, let scripts decide.
Why it matters for code review
Always-on factories FOMO when agents go quiet; the worse FOMO is a loop that keeps merging while spend, alerts, and improver thrash are invisible. Continuous verified momentum compounds when merge is a deterministic gate, task size is a cost lever, and failure modes upgrade the harness (credit monitor, human-routed alerts, cooldowns, escalation caps) instead of another heroic session. Pair with Cole’s Archon holdout perimeter when the factory must not rewrite its own rules, and with Guild’s separate reviewer when independence between write and decide is the scarce control.