OpenAI's single Dot vs Anthropic's 1,000-agent swarms
A comparison of two approaches to long-running agents: OpenAI gives one Dot with its own cloud computer, while Anthropic has a lead agent write a program and hire up to 1,000 sub-agents, catching 66 of 70 hidden bugs in testing.
OpenAI gave you one Dot. Anthropic gave you 1,000 agents. Two bets on long-running agents. OpenAI: one agent with its own cloud computer. Swarms come later. Anthropic: a lead agent writes a program and hires up to 1,000 agents per task. Live in the API now. Anthropic’s test: 70 bugs hidden in 116,000 lines of code. One agent, 3 attempts: 14 to 27. Multiagent workflow: 66. Every run. That’s 2.4x the single agent’s best attempt. What it looks like on a real project. Scenario: move a 400-module monorepo from JavaScript to strict TypeScript. > Phase 1. One agent maps dependencies and cuts the repo into 400 pieces. > Phase 2. 400 agents, one per module, rewrite the code. 64 run at once, about 7 waves. > Phase 3. 400 reviewer agents. Each checks someone else’s module and hunts for what the author missed. > Phase 4. Anything that fails review gets redone and rechecked. Budget up to 150 agents. > Phase 5. One agent runs the tests and writes the report: done, failed, uncovered. About 950 agents total. Under the 1,000-per-run cap, inside the 24-hour run lifetime. You sleep. The main thread stays free and reports progress by phase. The 66 out of 70 lives in phase 3. A second agent reads the code without the first one’s bias. The honest part: > every agent burns tokens, billed via the API, not your subscription; > “completed” means the program finished, not that everything passed. Read the report; > skip the session budget and this run eats your money before you wake up. Start with a smaller repo and tight limits. Then scale. I’d price it at 65¢: OpenAI shows multiple Dots on one task before March 1, 2027. Docs and templates in the reply.
This post quotes another post: x.com/vladic_eth/status/2078046195658662381