Anthropic just open-sourced ten finance-focused agent templates, backed by a $1.5 billion joint venture with Blackstone and Goldman Sachs. Instead of a slick demo, the GitHub repo reveals a reference architecture built for regulated workflows. You get one system prompt per agent, stored as a single markdown file that drives both interactive and headless modes—no runtime prompt assembly, just a Git-tracked source of truth. Every managed agent splits roles into subagents: only one can write, one reads documents, another enforces rules. That separation of duties mirrors bank workflows and limits blast radius if an LLM misbehaves. They lock down data flow with JSON schemas, not free-text strings. The KYC doc-reader, for example, returns a packet_id, entity, and UBO list, each field constrained by length and regex patterns. Prompt injections can’t masquerade as instructions. Every agent also starts with clear guardrails in its system prompt—“recommend, don’t decide,” “produce reports, don’t post to ledger.” Those instructions live in every run but stop short of hard enforcement. Three of the ten templates trigger the EU AI Act’s high-risk rules on August 2, so you’ll need risk registers, accuracy statements and other docs to fill the remaining 20 percent of compliance work.