Agents Can Reason. They Still Can't Search

(dipkumar.dev)

5 points | by askhn1234_12 2 hours ago

1 comments

  • frank1677 1 hour ago
    Good read

    The interesting part is that every layer independently reinvented retrieval: RAG for documents tool routing for MCP/actions skills/workflow loading context compaction/navigation for long-context agents

    Different abstractions, same core problem: “How does the agent find the right thing at the right time under cost/latency constraints?”

    Feels like search is becoming the actual runtime for agents.