Anthropic engineer: “80% of our engineers are using self‑improving loops. Now everyone is building agentic Graphs. In 4-6 months, we’ll all be building graphs to orchestrate self‑improving agents. No more prompting.” in a 20‑minute talk, Anthropic engineer explains how to build self‑improving agentic systems from scratch. Worth more than... a $500 agentic course. Watch this video, then read the article below on how to become a graph architect.show more

Am I missing something, the video doesn't talk about graphs at all

A graph is just a loop with moa. Everyone is mixing abstraction layers. Still a loop.

'graph architect' is sending me bro went from prompter to plumber in one linkedin post

I thought we where past graphs? You should try continuous fields.

"No more prompting" gets said every 6 weeks. What's actually happening is quieter: the prompt matters less, the verifier matters more, and the biggest unlock is verifiers from a different model family than the builder. That part nobody's selling a course on.

To be honest this methodology isn't really affordable for anyone other than anthropic employees. I can use my max subscription in like 30 minutes using workflows because all the subagents try to fill context

this is definitely a must-watch, thanks for dropping this alpha

You are welcome buddy 🫂

'you can think about Claude as a light source', could these guys be further up their own asses?

Can not stress how important it is to ignore as many of these ai influencers as possible in order to get the best results.

the actual bottleneck is gonna be verifying the graph's output, not building the graph itself

No one has convincingly explained how this is an actual improvement. What do you mean you don’t prompt. You need to tell ai what you want to do right? I end up having to steer Claude in the right direction constantly. You tell it to make a site and things are out of place, or too far to the left for example. How do loops actually solve this. It has no idea what you are imagining

Take a look at

is this why claude code is super slow?

the hard part will be debugging why the graph chose a path

Prompt engineering then context engineering then loop engineering then graph engineering. Wtf is next??

but who's auditing the self-improving part when it goes sideways feels like we're hyped about the engine before anyone built the brakes

ive learnt from movez more than from school

Thanks bro. Happy it’s useful for you

I’m so tired that I’m not even going to bother figuring out what “graphs” are. I’ll just wait until this gets rolled into the harnesses without custom dev work….

did this and open sourced it, look up nomos on hugging face, the main thing is for people to know the work exists instead of continually rebuilding the wheel

graphs are just the new prompt format - the edge is still in evals and debugging

@0xMovez sounds wild. gotta wonder how this changes the usual dev workflow though

Graph engineer doesn't work, but sure is good marketing.

Whats stopping these graphs from becoming too complex?

anthropic engineer: "80% of our engineers are using self-improving loops. in 4-6 months everyone will be building agentic graphs. no more prompting." the shift is already happening. from "write better prompts" to "build systems that improve themselves". most people are still stuck on the first part.

A graph is just a loop with moa. Everyone is mixing abstraction layers. Still a loop.

4-6 months is wild because that timeline keeps getting shorter every time someone says it we are genuinely not ready

systems outlive prompts

One thing I'm still unsure about: as these graph workflows get larger, where does most of the context overhead actually come from? Shared state, tool outputs, or subagents repeatedly gathering the same information?

being building graph since last year

He's selling tokens

yeah I mean, I said it but yes glad it's working now.

@summarize_that can you summarize this?

Sloppy