Databricks research: Multi-step agents outperform single-turn RAG when answers span databases and documents. Databricks shows multi-step agents beat single-turn RAG by 20%+ on hybrid tasks by decomposing queries and coordinating SQL and vector retrievals. If your systems must join documents and databases, design for decomposition and cross-tool coordination rather than one-shot retrievals (Principle 09, Principle 06).

Show HN: Kontext CLI – Credential broker for AI coding agents in Go. Kontext injects ephemeral, scoped credentials into AI coding agents and logs every tool call for enterprise-grade governance without storing keys. Adopt credential brokering and per-call audit logs as standard plumbing for agents that touch production systems (Principle 10, Principle 06).

Securing non-human identities: automated revocation, OAuth, and scoped permissions. Cloudflare adds automated token revocation, OAuth visibility, and resource-scoped RBAC to prevent credential-driven breaches for non-human identities. Treat short-lived tokens, visibility, and scoped RBAC as baseline agent security controls when exposing services to autonomous workflows.

Anthropic’s Claude Managed Agents gives enterprises a one-stop shop but raises vendor lock-in risk. Anthropic embeds orchestration in the model layer to speed deployments, but that convenience increases vendor lock-in and opaque control planes. Balance the deployment velocity of managed orchestration against portability, auditability, and your ability to insert independent gatekeepers (Principle 09, Principle 15).

Anthropic details AI agents accelerating alignment research through ‘weak-to-strong supervision’. Anthropic uses agentic pipelines where weaker models run experiments and supervise stronger models to speed alignment and evaluation workflows. Use agents to automate test generation and model checks, but instrument and audit those automated supervisors—automated supervision scales research, not trust by default (Principle 16, Principle 14).