Graph Engineering, a pattern for organizing AI agents, embeds self-review cycles into the architecture itself rather than the prompt. Anthropic’s implementation reports the tradeoff: a 90.2% improvement over single-agent Claude Opus 4, at a cost of roughly 15x the tokens.
Self-review isn’t new. Language models already improve when prompted to critique their own work. The news is quantification. You now know the precise tradeoff: is 90% better output worth 15x the cost? For a high-stakes decision or intricate analysis, probably yes. For routine questions, no.
Architecturally, the real innovation is separate reviewer nodes with isolated context windows and scoped tools. This prevents context pollution across parallel work while letting each agent specialize. It also means you can scale horizontal tasks without blowing out a single context window.
What’s missing: practical guidance on when this pays off. The eval doesn’t specify which problems they tested against or where the efficiency boundary lies. That distinction matters for real deployment decisions.