Wink Pings

Karpathy Says Prompts Will Disappear; What Really Remains Is the Graph

A tweet condenses the path from LLMs to Graphs into four steps, and someone in the comments adds a sharp correction: the agent loop is already a graph. The video and context are below.

In a one-hour lecture, Andrej Karpathy makes a judgment: prompts will fade, and what remains in the end is the graph. The tweet's author writes the path as LLMs → Prompts → Agents → Graphs and says most people are still polishing prompts, while the graph is the layer that will remain.

The graph here is not about drawing a flowchart. In an agent system, a graph organizes model calls, tools, state, conditional branches, and loops into nodes and edges. A single conversation is initiated by a prompt, while structure across multiple calls is maintained by the graph. As tools, branches, and failure recovery multiply, implicit control flow turns into an explicit graph.

In a reply, Jeff Bruchado points out that the arrow from agents to graphs skips a step: most agent loops are already graphs, with a node pointing to itself. This addition is crucial. A single agent's while loop is a degenerate graph; only with multiple tools, multiple roles, and conditional branches does the graph need to be expanded.

Karpathy's judgment is more like a shift in engineering focus. The prompt is the entry point for a single interaction, while the graph is where the system maintains state, constraints, and observability across multiple calls. Demos show only prompt techniques, but production still comes back to orchestration, state, and failure recovery.

The tweet's author says the lecture is free and worth more than most $500 graph-engineering courses. That claim has a promotional flavor; the video itself is worth watching, but you should make your own judgment. The original tweet is here.

发布时间: 2026-11-17 23:27