One prompt ran Codex for 17+ hours with 109 sub-agents and 2B tokens
andimarafioti · x · 2026-11-06
Developer andimarafioti shared an extreme agent run: a single Codex prompt autonomously worked for 17+ hours, consumed 2B+ tokens, and launched 109 sub-agents — stopping only when he asked it to check in regularly.
The setup (screenshot): a reviewer agent listed 88 improvement points across the codebase, then one sub-agent per task with a coordinator orchestrating — tasks ordered and grouped, sub-agents working in parallel without stepping on each other, reviewing each other's work, and producing mergeable PRs instead of 88 scattered ones. Most output was good, easy-to-merge work; the main issue was verbosity (thousands of doc lines per PR).
The work feeds his voice deployment for Reachy Minis, a vibe-coded project that served 3k+ distinct users in October across 45k conversations averaging 30 minutes. His takeaway: in February OpenClaw finishing one clearly defined goal felt surprising; now a single prompt pursues a bundle of loosely defined goals — his job is increasingly finding where he's the bottleneck and getting out of the way.
More from coding & agent
- ~780 AI slop accounts identified on X; users deploy agents to auto-block them daily — TheZachMueller · 2026-11-06
- Formal proofs of tens of millions of lines of code are now feasible, says Anthropic engineer — morqon · 2026-11-06
- Terminal-Bench's two broken tasks traced to Debian 11 archive repo migration — PawelHuryn · 2026-11-06
- Stable AI proposes DPDM: frozen LLM representations plus shared predictor boost agent decisions — stable-ai · 2026-11-06
- Reddit devs compare agent memory approaches: summaries, MEMORY.md, wikis and vector DBs all fail differently — ombre-ai · 2026-11-06
- MCP vs API vs CLI: build both on one operations layer, prioritize MCP — hottown · 2026-11-06