Anthropic ex-engineer runs an internal graph that costs $6 a month and catches what a $300,000 eval suite misses. No retrieval layer. Seven nodes and one rule about who's allowed to change their mind. He published the whole schema. His version starts from the opposite idea. A graph is... not an execution order. It's a memory of why. Seven nodes. Every edge carries the reason it exists: > INTENT - states what the task is for. Never how > DECOMPOSE - splits it into steps, each with a stated assumption > WORKER - executes one step. Sees nothing else > AUDIT - checks the output against the assumption, not the goal > DRIFT - compares the current step to INTENT and flags divergence > LEDGER - stores every decision with the assumption that justified it > ROOT - holds the graph, and when an assumption breaks, re-runs every step built on it Six nodes act. One node remembers why they acted. Every step carries the assumption that made it correct, so a false one only reruns what stood on it. That's the entire design. A pipeline that forgets its reasons has to redo all of it or trust all of it. He replayed a month of agent runs. 4,100 steps, 380 built on an assumption that was wrong by day three. The old pipeline shipped all 380 and linked none of them. Everyone else builds graphs where output moves forward and the reasoning evaporates. He built one where the reason travels with the result. The blast radius is the output nobody else produces. The article below is the full build - node prompts, the assumption format, the invalidation rule that finds every step downstream of a broken one. Save it. You'll want it open in the othershow more

@grok is this dev twitter ai slop grift or is there any papers on arxiv or something that actually back this kind of approach in a measured way with evals. i know for a fact the dashboard shown is slop but is there anything useful to take away from this

“A graph is a memory of why” is the primitive. Freeze the intent. Bind every claim to evidence. Record who may revise it. Then sign the handoff so another agent can continue without hidden context or inherited trust. We have that architecture. We’re building it. Get at me.

@grok can tou explain this whole graph thing for noobs?

One sub to rule em all?

$6 beating $300k, wild

@0xWast3 wild how sometimes the simple setups catch what the expensive ones miss. makes me rethink my own eval approach.

Tracking assumptions makes agents far more reliable.

Claude build me a cool graphic that does absolutely nothing but gets social media likes

a $6 graph beating a $300k eval suite is the whole lesson about knowing your actual failure modes. most eval spend is buying coverage of problems you don't have

how strict is the rule on who’s allowed to change their mind?

Your work is inspirational

root invalidation beats a $300k suite that never stored why

insane bro

Dope

Brilliant read. Thanks bro

Graph as memory of why, not just execution order. Assumption tracking is the real control plane.

nice animation

"Claude make an interactive dashboard about knowledge graph then write me a story to post on X"

This tracks with how I run things too. I'm more a planning, step by step, test water kind of guy. I do take risks, but they're measured ones, not all in leaps. A system that keeps its reasoning alive is measured risk built into architecture.

It's amazing how they manage to do that

这思路感觉很牛

Да схуяли она обходится 6 долларов?

noted this gem fam!

Fascinating an engineer can strip the evidence component of Atomic Genesis Decision Record attach a graph to it and really believe is their original work. Removing the institutional value to make it appear they have done something mea ningful. Bravo

Sezz béé3e

What lib is the ui built on top of?

memory matters

I've never seen such connection nodes, apparently the engineer came up with them himself

this is the most elegant thing i've read about agent design in months

she leaked it for us now

Memory drift happens because old context chunks keep winning the similarity lottery. You don't need a bigger context window, you need clean invalidation rules. Been testing Atomic Strata Memory to force strict state updates instead of just blindly appending.

这个思路有意思。我一直觉得eval套件堆钱不如把结构搞对。七个节点和一条规则，这种精简反而可能更可靠。有没有可能开源出来？想看看schema。

that's actually brilliant building a reason graph is a game changer for agents