{
  "id": 20414241,
  "title": "Claude Code plans, Codex builds: what 152 task cards taught me about governing two coding agents",
  "url": "https://urgent.news/2026/11/17/claude-code-plans-codex-builds-what-152-task-cards-taught-me-about",
  "topic": "ai",
  "section": "AI",
  "published": "2026-11-17T10:04:08.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/tomxiaoyz/claude-code-plans-codex-builds-what-152-task-cards-taught-me-about-governing-two-coding-agents-2c2k"
  },
  "original_language": "en",
  "account": "Two coding agents, Claude Code and Codex, were used to manage personal repositories over a two-week period. The setup involved 152 task cards distributed across seven repositories. Initially, Claude Code handled planning and writing tasks, while Codex implemented the code and conducted tests. As the process evolved, several key factors contributed to increased efficiency and reliability.\n\n1. Handoff in the repository: Instead of relying on chat conversations, handoffs were made through Markdown cards in the task branch. This approach provided a clear record of objectives, acceptance criteria, boundaries, authorization parameters, tests, and logs, ensuring that everyone involved could easily review and track progress.\n\n2. Card size and round limits: To prevent overly complex tasks, cards were capped at 12 KB in size and allowed only two rounds of execution. If the task required more than two rounds, it was split or escalated to a human for review. This helped maintain manageable workloads and reduced the chances of errors due to environmental issues.\n\n3. Cost analysis: The cost of executing tasks was measured by the number of tokens used by Claude Code and Codex. Claude Code, responsible for planning and reviewing tasks, had a significantly higher token volume compared to Codex, which focused on implementation. This ratio highlighted the importance of efficient planning and reviewing in reducing overall costs.\n\n4. Audit and gate: Introducing a turn-end hook that audited the diff for budget overruns, unauthorized files, and other issues allowed for early detection of problems. When the audit failed, Codex was sent back within the same turn, providing an opportunity to correct the issue before escalating to a human reviewer. This helped minimize costly revisions and improved the overall reliability of the system.",
  "summary": "I maintain several personal repositories: a data platform, a quant research repo with hard \"never touch live trading\" lines, a few e-commerce tools, and a content pipeline. Since late October I have split the work between two agents: Claude Code defines and accepts. It decides what to build, writes the task, and checks the result. Codex implements. It writes code and runs tests inside the spec it…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}