{
  "id": 3388288,
  "title": "Loops vs Graphs: Why Agent Architecture Needs Both (and a Compiler Between Them)",
  "url": "https://urgent.news/2026/08/26/loops-vs-graphs-why-agent-architecture-needs-both-and-a-compiler",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-26T00:09:51.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/bozoinc/loops-vs-graphs-why-agent-architecture-needs-both-and-a-compiler-between-them-14n0"
  },
  "original_language": "en",
  "account": "The debate in the agent ecosystem is between two approaches: Camp Loops, which sees agents as iterative control structures, and Camp Graphs, which views agents as interconnected nodes in a graph. Both perspectives are valid but neither fully captures the complexities of agent architecture.\n\nThe article argues that neither camp addresses the critical need for verification of the capabilities that agents use. Currently, agents are wired together from unverified repositories, with no guarantees about the quality or safety of the code they depend on. This leaves loops vulnerable to running flawed or malicious capabilities, and graphs susceptible to supply chain attacks.\n\nThe proposed solution is HURCULES, a compiler that sits between repositories and agent runtimes. HURCULES takes raw GitHub repositories and produces verified capability packages that can be used by either loops or graphs. HURCULES performs several key functions:\n\n1. It creates a deterministic map of the repository's structure and dependencies.\n2. It analyzes the code with an LLM and produces a set of candidate capabilities, each linked to the original files.\n3. These capabilities are challenged by a devil's advocate to ensure they are legitimate and safe.\n4. Human approval is required before the capabilities are compiled into a verified package that includes a capability spec, file-level evidence, provenance information, approval trail, license classification, and security scans.\n\nFor loops, HURCULES provides verified nodes to plan and act upon, ensuring the loop runs on legitimate capabilities. For graphs, HURCULES supplies verified edges, allowing nodes to communicate with attested capabilities rather than raw LLM-generated outputs. This verification adds an essential layer of trust and accountability to agent architecture.\n\nThe article concludes that the agent ecosystem needs a verified capability supply chain, not just more specialized frameworks. HURCULES fills this role by compiling a registry of verified capabilities from any repository. Loops and graphs are the consumers of this verified supply chain, while HURCULES acts as the trusted intermediary. The author encourages adopting HURCULES to build more reliable and secure agent systems.",
  "summary": "The False Dichotomy The agent ecosystem is split into two camps: Camp Loops (Boris Cherny, OpenAI Agents SDK, LangGraph): > \"Agents are loops. Plan → act → observe → repeat. The loop is the atomic unit.\" Camp Graphs (Steve Yegge, Gas Town, LangGraph DAGs, CrewAI): > \"Agents are graphs. Nodes are agents/tools. Edges are handoffs. The graph is the architecture.\" Both are right. Both are incomplete.…",
  "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."
}