Nine agents ran for 8 hours, burned 500K Opus 5 tokens, and shipped 18 PRs
HaktanSuren · x · 2026-07-27
The author says they are literally watching agents argue with each other after letting a swarm run overnight.
In the quoted post, they report 9 agents running for 8 hours, burning half a million Opus 5 tokens, and producing 12 backend PRs and 6 frontend/iOS PRs. Most of the work drifted into ops, performance, and security even though the ask was for new features. The takeaway: orchestration quality may matter more than raw model cost if you want paid-model agent swarms to be useful.
More from coding & agent
- Opus 5 is tested with Roblox MCP in a cute tool-use demo — JasonBotterill · 2026-07-27
- Codex powered the whole project, including the Cloudflare agent setup — jeremyjordan · 2026-07-27
- A ChatGPT Work planner uses a custom MCP server, Cloudflare state, and an e-ink display — jeremyjordan · 2026-07-27
- AI data science portfolios need two real projects, not ten toy demos — mdancho84 · 2026-07-27
- A legal-document agent stack uses Google ADK, MCP, and four parsers to cut hallucinations to zero — MaryamMiradi · 2026-07-27
- LLMs make vibe coding work — but they also add layers of weird complexity — StewartalsopIII · 2026-07-27