Anthropic Ex-Engineer: $6 Graph Catches What $300k Eval Misses
AdiPolak · x · 2026-08-25
An ex-Anthropic engineer shared an internal graph architecture costing $6/month that catches issues missed by a $300,000 evaluation suite. It uses no retrieval layer, consisting of just seven nodes and one rule.
The core concept: A graph is not an execution order but a memory of why things happened. The 7 nodes are:
- INTENT: States what the task is for, not how.
- DECOMPOSE: Splits into steps with stated assumptions.
- WORKER: Executes one step, seeing nothing else.
- AUDIT: Checks output against the assumption, not the goal.
- DRIFT: Compares current step to INTENT and flags divergence.
- LEDGER: Stores every decision with its justifying assumption.
- ROOT: Holds the graph; re-runs steps built on broken assumptions.
The author published the full schema, arguing that multi-agent systems in 2026 shouldn't rely on slow, sequential straight-line processes.
More from coding & agent
- AI Agents Need Four Types of Memory to Mimic Human Capabilities — _jaydeepkarale · 2026-08-25
- Postmortem of a Hindi-English voice agent for fintech: number readback and real concurrency were the real problems — admrys · 2026-08-25
- The future is harness-independent and LLM-independent: SaaS giving agents instead of MCPs shows narcissism — shensi · 2026-08-25
- Balance Speed and Understanding When Using AI Coding Agents — arpit_bhayani · 2026-08-25
- Tencent releases GameXpert-Bench to evaluate coding agents in game development — Tencent-Hunyuan · 2026-08-25
- Grok Bot Reads Order History to Build Perfect Shopping Cart — elonmusk · 2026-08-25