{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreictg4ltbyqlcwgnrlfflsfyt6w3tzmr5zwupal6cb7g5b5wmbk6z4",
    "uri": "at://did:plc:lk3jfj3zq4k4wxnk474axylu/app.bsky.feed.post/3mnvdf5mjhzt2"
  },
  "path": "/t/codex-rate-limits-discussion-thread/1378553?page=19#post_381",
  "publishedAt": "2026-06-09T22:48:54.000Z",
  "site": "https://community.openai.com",
  "textContent": "So-called artificial intelligence.\n\nAt this point, the AI should be smart enough to decide which tools it needs, when it needs them, and whether they should be enabled or disabled. Right now, it can decide when to use a tool, but the loading and management of those tools is still mostly manual.\n\nThen what exactly is “intelligent” about this setup?\n\nThe ironic part is that skills are supposed to make the agent more efficient and reduce token usage. But in practice, if you load too many of them, they seem to create the opposite problem. So now the user has to carefully pick only a small number of useful skills, remove the unnecessary ones, and be even more careful with MCP tools.\n\nThat feels backwards. The AI should manage its own toolbox. I should not have to act like a system administrator just to stop my “artificial intelligence” from wasting tokens.\n\nHopefully, future models will handle this automatically with a smarter loading system, where only the truly relevant tools are pulled in when needed, without bloating the context.",
  "title": "Codex Rate Limits Discussion Thread"
}