Is Prompting Over? First, Look at the Bill for 300 Agents

An X post puts “Prompting is basically over” right at the start. The post relays a claim from someone labeled as an Anthropic Research Lead: 99% of engineers are already running more than 300 self-improving agents, and everyone is moving toward graphs. In a 20-minute video, the path is compressed into five steps: One Agent → Loops → Plan Mode → Graphs → Self-Improving System.

Let’s break down the terms first. Here, graph means an orchestration graph, not data visualization. It treats agents as nodes and uses edges to define calls, fallbacks, parallelism, and checkpoints. When an agent gets stuck in a loop, the graph decides who gets the next step. The post’s claim is that prompting is the old workflow, and the graph outside the model is the new workflow. Graph orchestration is not a new term in engineering. Libraries like LangGraph have long turned nodes and edges into a product interface. What is new is the scale described by “300+ self-improving agents.”

In the quoted post, m4lka offers a more concrete summary: “Your agent is not autonomous. You are its loop.” Your agent is not autonomous; you are its loop. You check step by step because there is nothing else in the system that can check. He wrote about seven levels from babysitting agents to running a fleet. This observation is closer to everyday reality than “prompting is over”: much of the automation in many agents actually depends on humans filling the gaps.

The comments did not simply go along with the conclusion. Some asked about state consistency: if a node changes the shared context during a run, what does the graph do—does it use a DAG or a cyclic graph? Some asked about money: at current token costs, who outside the frontier labs can afford it? Others were openly sarcastic: AI labs charge by token, and now they tell you one agent is not stable enough, so you should spin up a group of agents to manage one another. This comment, from Ariel Jalali, makes its point clearly.

So is the video worth watching? If you care about agent orchestration, the five-step path can serve as an index. But “99%” comes without a sample or definition. 300 agents do not equal 300 employees, and the bill will arrive first. Graphs can solve orchestration, but they do not solve model capability. Without state management, evaluation, and rollback, a graph is just a more expensive loop. For developers: start with a small graph, add versioning or locks to shared context, and log the inputs and outputs of each node. For ordinary users: do not panic when you see “over.” Prompting has not disappeared. It has become the input design for every node in the graph.

Original post: https://x.com/fomoebka/status/2115837233869554052

m4lka’s quoted post: https://x.com/m4lkxx/status/2114338456750997922