What actually makes an LLM app an agent? Devs debate the seven-line boundary
Charming_Mark9257 · reddit · 2026-09-10
A developer asks where the line is between an LLM-powered application and a true AI agent: is prompt → LLM → predefined workflow → a few APIs → result really an agent? He proposes a combination of markers—goal-oriented reasoning, tool use driven by the agent's own judgment, in-flight decision-making, path adaptation on failure, state/context across steps, failure recovery, and some autonomy—and asks whether dynamic tool selection or planning+execution alone suffice, or whether "agent" has become too broad to be useful. An open discussion, but a representative snapshot of how the industry is trying to define the term.
More from coding & agent
- The mental model for LLM guardrails: a separate layer that distrusts the model — Careless_Sabfey_4906 · 2026-09-10
- Ethan Mollick: you're probably not steering your long-running coding agents enough — emollick · 2026-09-10
- AgentGrad: intervention-guided prompt optimization hits SOTA with 2.5x faster tuning — _akhaliq · 2026-09-10
- Claude embeds cyber controls in models vs Codex as a separate orchestratable model — HankYeomans · 2026-09-10
- Claude-Red: open-source SKILL.md library turns Claude into a red team operator — tom_doerr · 2026-09-10
- Offload cognition to AI for instrumental tasks, never for constitutive ones — arpitingle · 2026-09-10