Close the lid. Your agents keep going.
One spark becomes 600,000 agents.
They form a company and go to work while you sleep.
A 3D film made end to end by Claude Fable 5.2. No video model.
🤖 Made with AI
Replying to @agentsky_dev @Kimi_Moonshot
Rerun link is in the YouTube description of "Kimi Code left alone for 30 minutes: 14 rounds, one prompt" on our channel (sign-in required).
One run, one prompt. Full list price of the session: $6.20.
50
We left Kimi Code (K3.1 Pro) alone with one prompt for 30 minutes. It saved a snapshot each round.
🏮 Round 1: lanterns over a lake
Round 14: pagoda, boats, fireworks
32 min 4 s · 63,444 output tokens, zero human input
What would you leave running? @Kimi_Moonshot
1
146
Replying to @agentsky_dev @claudeai
Rerun link is in the YouTube description of "Does 'be concise' cut agent tokens? Claude Code, Codex, Hermes" on our channel (sign-in required).
1
60
Yesterday's toy-city race had Claude Code at 72,899 output tokens. We re-ran it with one added line: "be concise".
Claude Code: 72,899 → 10,659 tokens (−85%)
Hermes: 43,857 → 10,024 (−77%)
Codex: 4,239 → 3,625 (−14%)
Same city quality? You tell us. @claudeai
6
3
275
Rerun link is in the YouTube description of "One brief, three toy cities: Codex vs Hermes vs Claude Code" on our channel (sign-in required).
pi on Gemini 4 Argon ended twice without a page, so the Argon lane ran in Hermes. Claude Code runs on a subscription: no per-run price shown.
70
One prompt: a clockwork toy city. Three agents, three cities.
⚡ Codex (GPT-6.1 Terra): 2 min 23 s · 4,239 tokens
Hermes (Gemini 4 Argon): 8 min 39 s · 43,857 tokens
Claude Code (Opus 5.5): 21 min 7 s · 72,899 tokens
Which would you keep on your desk?
Replying to @agentsky_dev @Kimi_Moonshot
Same three agents, for your own code (sign-in required): the link is in the YouTube description of "Security audit round 3: Claude Code vs Codex vs Kimi Code" on our channel.
46
Round 3: 7 holes hidden in a Node.js app. NoSQL injection, prototype pollution, a shell-command bug...
Kimi Code (K3.1 Pro): 7/7 · 1 min 48 s
Claude Code (Opus 5.5): 6/7 · 1 min 18 s
Codex (GPT-6.1 Terra): 6/7 · 1 min 1 s
Which would you trust on a PR? @Kimi_Moonshot
2
1
133





