Agents, GraphRAG, open models, and cheaper frontier inference

Welcome to Agents Week — Cloudflare kicks off Agents Week to define an agent-native Agent Cloud, focusing on primitives, developer lifecycle, security, and the agentic web. This lays out operational and security primitives you’ll need when building agent infrastructure — treat it as a playbook for agent orchestration and lifecycle (Principles 09, 15).

Ron van Kemenade, Group COO, Lloyds, on agents, COBOL, automating fraud detection — Lloyds uses agentic AI to analyze and refactor COBOL, validating replacements in parallel before decommissioning legacy systems. It’s a concrete pattern for using agents as discovery-and-validation lanes during modernization, showing how to avoid single-player rollouts and keep production safe (Principles 03, 16).

Stop graphing everything: When GraphRAG actually beats vector RAG — GraphRAG improves corpus-level reasoning over vector RAG for global questions but imposes extra indexing and compute costs. Outcome engineers should weigh GraphRAG’s stronger global reasoning against its operational overhead and choose vector vs graph based on query patterns, latency, and cost (Principles 06, 11).

Running Kimi K3 on MI355X at Better Performance per Dollar Than B300 — MI355X runs Kimi K3 with superior performance-per-dollar compared to B300/B200, enabling cost-effective frontier-model inference. If your agent pipelines require heavy on-prem or edge inference, this shifts deployment and SLA trade-offs and can materially lower running costs (Principles 12, 16).

Latest open artifacts (#23): Laguna S2.1, Inkling, & Kimi K3 show the utility of open models on the Pareto frontier — New open-weight releases like Inkling, Hy3, and Laguna-S-2.1 demonstrate that open models can sit on the Pareto frontier with proprietary alternatives. That changes sourcing decisions: you can run stronger models in-house, customize toolchains, and ship verifiable artifacts faster, but you must also revisit containment and governance (Principles 08, 10).