Anthropic engineer: "You don't need better prompts. You need graph engineering that makes agents remember everything." In 25 minutes he shows how engineers at Anthropic wire agents into a graph where each has a job, they run in parallel, verify each other, and share a memory that never resets.... This beats any paid course on building agents I've seen. Watch it, then read the full guide on graph engineering below.show more

Original video:

The parallel verification piece is what most people skip entirely.

That´s it

“graph engineering” is just workflows. It’s nothing new, hell lang graph has been doing it a while. It’s the same thing Anthropic argued against with their “thin harness” thing, yet now they’ve completely changed their minds?

This is painful to listen to. I wonder why people who do not know how to properly talk participate in these conferences. He probably has some idea on a topic, but his speech is incomprehensible with all those parasitic words "like," "you know," and others.

this is the infrastructure shift that matters

The graph memory architecture makes way more sense than trying to cram everything into a single prompt context window.

I feel like we’re already doing some of this through artifacts and Markdown-based diagrams. What really interests me is how GraphRAG could extend that into richer, more structured visual reasoning.

That video has nothing about graphs

shared memory changes everything

The future of AI agents is shifting from better prompts to better architectures — graph-based systems enable specialized agents to collaborate, share memory, verify outputs, and solve complex tasks autonomously....

Graph Engineering builds agency and control over executions I built GraphARC to explore that idea: let agents dynamically build their own execution graphs, while deterministic tool controls the laws they can't break. Laws are defined by usrs. Test it out!

in theory, that sounds great but these are unintelligent ghosts that compound error and discontinuity, which is favorable in creative work, just not all works

That is very nice of you to not start by "instead of spending Xmin on Netflix tonight..."

Parallel agents with shared memory sounds clean until two agents write conflicting updates to the same memory node at the same time and you have no merge policy. Has anyone at Anthropic addressed how graph agents handle concurrent write conflicts without human arbitration?

So... a database?! Graphs are UIs for developers who can't CLI 😎

Doesn't work in reality. It might look cool in a 5 minute tech talk

graph engineering is the real unlock memory that never resets means agents can build on past context

This is how AI starts to scale

That's a lot of "like"

We haven't earned our valuations. Facts.

Both are useful for communicating ideas. Human communication start with speaking, the prompts. Graph help organize our ideas when sentences not enough to match what’s in our brain. Real problem is not which one better, there’s no easy/quick tools to turning our ideas into graph.

Ohh! Is it really that incredibly smart?

25 minutes of graph engineering explained clearly. a perfect place to start diving into the topic

Like

The video and your content have 0 relevance!! Quite misleading!

provides unified scheduling for major global AI models, with particularly cost-effective and high-performing Chinese models.

I like this viewpoint. Progress starts with consistent effort.

shared memory is the missing piece; without it, most agent systems are just smart workers waking up with amnesia every morning.

グラフエンジニアリングって具体的に何をするの？気になる！

the future of agents is built on memory, not just prompts

Memory without orchestration becomes context hoarding. The real advantage comes from agents with clear roles, verification loops, and measurable outcomes.

The interesting part is “Anthropic engineer: "You don't need better...” but the proof is still missing.

The "memory that never resets" part is the thing worth watching closely. Persistent shared state across agents sounds clean until you get compounding hallucination, where agent B confidently builds on agent A's wrong assumption and neither flags it.

Graph engineering makes a lot more sense once agents stop being isolated chat windows. The moment multiple agents share context, roles, memory, and verification, the architecture becomes the product.

The future of AI isn't better prompts—it's better systems. Graph engineering is the real unlock. 🚀