Ask HN: AI Agents vs. Gateways vs. Harnesses
TL;DR
A Hacker News thread calls out the messy terminology in the AI agents ecosystem and proposes a cleaner taxonomy. The author suggests three layers: Harnesses (UI + system prompts + tools wrapped around an LLM, e.g. Claude Code, Gemini CLI), Gateways (connectors to communication platforms like WhatsApp or Slack), and Sandboxes (isolated, auditable runtime environments). The core complaint: almost everything gets labeled 'Agent', even products that only implement one narrow slice of the stack.
Nauti's Take
The thread is symptomatic of a broader problem: the AI industry names things for marketing value, not technical function, then acts surprised when nobody can orient themselves. The proposed three-way split – Harness / Gateway / Sandbox – is pragmatic and immediately useful, even if still rough around the edges.
What it misses is a dedicated category for true orchestration logic: the layer that coordinates multiple agents, manages state, and handles failures. Frameworks like LangGraph or Temporal belong there but barely get a mention.
Until the industry converges on shared vocabulary, the only fix is doing your own layer-by-layer audit before buying into any 'agent' platform.
Briefingshow
The lack of shared vocabulary wastes engineering hours every day – evaluating an 'agent framework' might land you a thin prompt wrapper or a full orchestration runtime, with no way to tell from the label. Clear layer definitions are a prerequisite for meaningful security models, interoperability standards, and honest market comparisons. Until 'agent' stops meaning everything, purchasing decisions remain a gamble.
Sources
