I recently moved from that method of work in Cline, where I ...

Rod

npub1r0d8u8mnj6769500nypnm28a9hpk9qg8jr0ehe30tygr3wuhcnvs4rfsft

hex

cd30c6ca21911c6bb630695afe1f40f4c6f3fafb7e9410e234527efd90915d96

nevent

nevent1qqsv6vxxegsez8rtkccxjkh7raq0f3hnltaha9qsug69ylhajzg4m9sprpmhxue69uhhyetvv9ujuem4d36kwatvw5hx6mm9qgsphkn7raeed0dz68hejqea4r7jmsmzsyrephumuch4jypchwtufkgznwagk

Kind-1 (TextNote)

2026-01-25T21:09:33Z

↳ 回复 Leo Wandersleb (npub1gm7tuvr9atc6u7q3gevjfeyfyvmrlul4y67k7u7hcxztz67ceexs078rf6)

With LLMs, being reckless wins. If you want to be a human in the loop, you won't spin up 50 agents to to work on 20 problems at once. How can we trus...

I recently moved from that method of work in Cline, where I experienced all the same issues, to running multi agents on auto-approve in Vibe-Kanban on top of Codex and Claude Code. This new method is much faster and the agents much smarter.

I think Cline's system gets in the way a lot.

IMO a system where you can trust their work mostly comes down to two things.

  1. Isolation. Each agent works in a discrete work tree and can make its changes only there.

  2. Verification. Giving the AI the tools they need to test their own work at multiple levels. Run TDD. Build and run your own Jest suite. Call the actual API endpoints and use them. Use Playwright and Maestro to capture screenshots before and after, and diff them. Etc

I think isolation and verification probably solve a lot of what you are concerned for.

原始 JSON

{
  "kind": 1,
  "id": "cd30c6ca21911c6bb630695afe1f40f4c6f3fafb7e9410e234527efd90915d96",
  "pubkey": "1bda7e1f7396bda2d1ef99033da8fd2dc362810790df9be62f591038bb97c4d9",
  "created_at": 1769375373,
  "tags": [
    [
      "alt",
      "A short note: I recently moved from that method of work in Cline..."
    ],
    [
      "e",
      "3c75f790160e6bd521944b9e660c290182c70c27d7ce86ad38f98f483313c39d",
      "wss://nostr.wine/",
      "root",
      "46fcbe3065eaf1ae7811465924e48923363ff3f526bd6f73d7c184b16bd8ce4d"
    ],
    [
      "p",
      "46fcbe3065eaf1ae7811465924e48923363ff3f526bd6f73d7c184b16bd8ce4d",
      "wss://nostr.wine/"
    ]
  ],
  "content": "I recently moved from that method of work in Cline, where I experienced all the same issues, to running multi agents on auto-approve in Vibe-Kanban on top of Codex and Claude Code. This new method is much faster and the agents much smarter. \n\nI think Cline's system gets in the way a lot.\n\nIMO a system where you can trust their work mostly comes down to two things. \n\n1. Isolation. Each agent works in a discrete work tree and can make its changes only there. \n\n2. Verification. Giving the AI the tools they need to test their own work at multiple levels. Run TDD. Build and run your own Jest suite. Call the actual API endpoints and use them. Use Playwright and Maestro to capture screenshots before and after, and diff them. Etc\n\nI think isolation and verification probably solve a lot of what you are concerned for.",
  "sig": "9446e30bcffd766f4ef15777d7d4ac279573bd13542b209db9866d22c2afc6174d7cfe8ac68cc943c3c53294f41329d282d469b478cf4422457676eb8e1bc89f"
}