Ex-Anthropic engineer's $6/month AI graph catches failures pricey evals miss

Aiden_Tech_Ai · x · 2026-08-28

A former Anthropic engineer built an AI graph costing just $6/month that catches failures expensive evaluation systems miss — with no retrieval layer.

At its core is a 7-node memory system for reasoning rather than an execution pipeline: INTENT defines the task's purpose; DECOMPOSE breaks it into steps and records assumptions; WORKER executes one step without seeing the rest; AUDIT checks output against its assumption; RIFT compares work against INTENT; LEDGER logs every decision and why; ROOT tracks dependencies and reruns affected steps when assumptions break.

The key idea: every step carries the assumption that made it valid, so when an assumption breaks the graph can locate and rerun only what's affected — which is why it's cheap yet catches errors others miss.

Original post →

More from coding & agent

coding & agent channel →