{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreigsin2olssnx4gxjvn4c7pcepvby2c5esiezubm6r4kfcjvoak2wi",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mji4d26l7w72"
  },
  "path": "/t/it-s-the-architecture-stupid-why-prompt-engineering-won-t-fix-agents/175246#post_1",
  "publishedAt": "2026-04-14T19:04:24.000Z",
  "site": "https://discuss.huggingface.co",
  "tags": [
    "Beyond Prompting: Decoupling Cognition from Execution in LLM-based Agents through the ORCA Framework",
    "GitHub - gfernandf/agent-skills: Agents should execute whenever possible — runtime for composable AI agent skills · GitHub"
  ],
  "textContent": "Borrowing from the classic “it’s the economy, stupid” — the same applies here.\nWe’re blaming prompts for what is fundamentally an architectural problem.\n\nPaper: Beyond Prompting: Decoupling Cognition from Execution in LLM-based Agents through the ORCA Framework\nCode: GitHub - gfernandf/agent-skills: Agents should execute whenever possible — runtime for composable AI agent skills · GitHub\n\n* * *\n\n* * *\n\nWe keep pretending that better prompts will fix LLM agents.\n\nThey won’t.\n\nWe’ve built an entire ecosystem of tooling, courses, and “best practices” around prompt engineering — as if the problem were linguistic.\n\nIt’s not.\n\nIt’s architectural.\n\n* * *\n\n## The uncomfortable truth\n\nLet’s be honest about what most agent systems are doing today:\n\n  * Take a task\n  * Generate a prompt\n  * Call the model\n  * Hope it “reasons” correctly\n  * Repeat\n\n\n\nThis is not a system.\n\nThis is recomputation disguised as intelligence.\n\n* * *\n\n## We are replaying cognition, not building it\n\nEvery time your agent runs, it:\n\n  * Reconstructs context\n  * Rebuilds reasoning\n  * Re-derives intermediate steps\n\n\n\nThere is no reuse of cognition.\n\nNo structure.\nNo persistence.\nNo abstraction layer.\n\nJust prompts.\n\n> We are not building systems. We are replaying thoughts.\n\n* * *\n\n## Why prompt engineering feels like it works (until it doesn’t)\n\nPrompt engineering gives the illusion of control:\n\n  * Add more instructions\n  * Add more examples\n  * Add more constraints\n\n\n\nAnd yes — performance improves.\n\nUntil it plateaus.\n\nBecause everything still lives inside a single forward pass:\n\n  * no memory of reasoning\n  * no composability\n  * no reuse\n\n\n\nIt’s like trying to fix software architecture by writing better comments.\n\n* * *\n\n## The real problem is architectural\n\nThe core issue is simple:\n\n> We are using LLMs as stateless reasoning engines.\n\nAnd then compensating for that with increasingly complex prompts.\n\nInstead of:\n\n  * modeling cognition\n  * structuring reasoning\n  * reusing intermediate steps\n\n\n\nWe regenerate everything every time.\n\nThat doesn’t scale.\n\nNot in cost.\nNot in latency.\nNot in reliability.\n\n* * *\n\n## What’s actually missing\n\nWhat’s missing is not a better prompt.\n\nIt’s a runtime layer that:\n\n  * encodes reusable cognitive steps\n  * separates reasoning into structured components\n  * allows composition instead of regeneration\n\n\n\nIn other words:\n\n> a system that reuses cognition instead of recomputing it.\n\n* * *\n\n## From prompts to skills (and where ORCA fits)\n\nInstead of:\n\n→ Prompt → Model → Output\n\nYou need:\n\n→ Skill → Execution → Structured Output\n\nNot conceptually. Operationally.\n\nThis is exactly what ORCA implements: a runtime layer where “skills” are reusable cognitive units — not prompts.\n\n  * defined inputs\n  * structured outputs\n  * explicit execution\n\n\n\nNo recomputation. No guesswork.\n\n* * *\n\n## Why most agent frameworks hit a wall\n\nMost “agent frameworks” today are:\n\n  * prompt orchestration layers\n  * tool wrappers\n  * retry loops with better formatting\n\n\n\nThey don’t model cognition.\n\nThey orchestrate prompts.\n\nThat’s not a runtime.\n\n* * *\n\n## The shift we actually need\n\nThe shift is not better prompting.\n\nIt’s architectural.\n\nFrom:\n\n  * stateless generation\n\n\n\nTo:\n\n  * structured, reusable cognition\n\n\n\nThat’s the gap ORCA is designed to close.\n\n* * *\n\n## Final thought\n\nPrompt engineering isn’t useless.\n\nIt’s just solving the wrong problem.\n\nWe’ve been optimizing the interface instead of the system.\n\nAnd it shows.\n\n* * *\n\nIf you’ve pushed prompt engineering far enough, you’ve seen the limit.\n\nThe question is:\n\nare you ready to try what replaces it?",
  "title": "“It’s the Architecture, Stupid” — Why Prompt Engineering Won’t Fix Agents"
}