Development & Automation

"The Creator of Claude Code Doesn't Write Prompts Anymore": He Builds Graphs Instead.

A quote is spreading that Claude Code creator Boris Cherny builds agents with "graphs" and "loops" instead of prompts. We check the facts and lay out what developers can do now.

· Huuush

"The Creator of Claude Code Doesn't Write Prompts Anymore": He Builds Graphs Instead

Key takeaways

  1. 1A quote is spreading on X that Boris Cherny of Anthropic, who built Claude Code, said "I no longer prompt agents; I build graphs and loops."
  2. 2He reportedly said that within 3 to 6 months everyone will be doing "graph engineering" for agent systems, which means designing how multiple agents connect and depend on each other beyond individual loops.
  3. 3The exact wording, the 3-to-6-month timeline, and the "10-minute talk" setting come only from a secondhand X post and are unverified.
Image expressing the idea that Claude Code creator Boris Cherny designs agents with graphs and loops rather than prompts: a network diagram of connected AI agent nodes
A still from the video attached to an X post quoting Boris Cherny (source: X @0xCodez, original X post)

A quote is spreading on X (Twitter) that Boris Cherny of Anthropic, who built Claude Code, said "I no longer prompt agents." According to the quote post by @0xCodez, he said in a 10-minute talk, "I'm building 'graphs' and 'loops,' and they build my agents for me. Within 3 to 6 months, all of us will be doing graph engineering for agent systems." The post drew 330,000 views and 1,500 likes and became a talking point in the developer community.

Loops Instead of Prompts, and Now Graphs

That Cherny says "I don't prompt Claude directly, I build loops that prompt" is not new in itself. From around June 2026, the remark spread among developers under the term "loop engineering," and he is counted among those who spread the concept together with Addy Osmani and others. What this X post conveys is the next stage: the discussion is moving beyond the loops of individual agents to "graph engineering," designing how multiple agents connect to and depend on each other.

After Prompts, Graphs?

As cases of running dozens or hundreds of agents at once grow beyond a single agent finishing work in a repeating loop, a diagnosis is emerging that the structure connecting them has become the bottleneck. In fact, Cherny has said that at Anthropic, "there are days I manage tens of thousands of agents." It ties into the trend that designing an agent org chart (who takes over what, and when) has become more important than prompting individual agents well.

Start by Drawing Nodes and Verification Loops

If this trend is real, the implication for practice is clear. The time you spent polishing a single prompt should shift to designing dependencies between agents and retry and branching logic for failures. For now, rather than moving on the "standard within 3 to 6 months" timeline as stated, a gradual approach is safer: first try small-scale loop engineering experiments already under way (iterative goal setting, then automatic verification, then retry), and consider a graph structure only when you need to tie several agents together.

It's Not an Official Term

The key phrase in this quote ("graph engineering becomes the standard within 3 to 6 months") and the "10-minute talk" setting come only from the secondhand X quote post. The original X post could not be opened directly because of a login wall, and neither Fortune nor Turning Post, which I cross-checked, confirmed this exact wording or when the remark was made (which conference). The gist of "building loops instead of prompts" and the spread of the term graph engineering itself are backed by multiple outlets, but the exact wording this post quotes and the "3 to 6 months" timeline are unverified, so it is best not to take them as definitive.

Frequently asked questions

What is the "graph engineering" Boris Cherny talked about?
It refers to designing how multiple agents connect to and depend on each other, beyond the repeating loop of a single agent. He reportedly said everyone will be doing this within 3 to 6 months.
How is it different from "loop engineering"?
Loop engineering is the idea of building a loop that prompts rather than prompting Claude directly, and it spread from around June 2026. Graph engineering is the next stage: a discussion about designing the structure that connects multiple agents.
Is this statement confirmed fact?
Multiple outlets back both the gist of "building loops instead of prompts" and the spread of the term graph engineering, but the exact wording this post quotes and the "3 to 6 months" timeline come only from a secondhand X post and are unverified.
How should I approach it in practice?
Rather than taking the "standard in 3 to 6 months" timeline at face value, a gradual approach is safer: first try small-scale loop engineering experiments such as iterative goal setting, automatic verification, and retries, and consider a graph structure only when you need to tie several agents together.

Related articles