Yanli Liu clones Anthropic’s open-sourced finance agent repo and identifies four key architectural patterns—single-source dual-delivery, write-holder isolation, schema-validated trust boundaries, and prompt-based guardrails—that set regulated-industry agents apart from generic frameworks. She also flags a 20% gap of non-code compliance artifacts (risk registers, accuracy declarations, transparency docs) needed to meet the EU AI Act’s high-risk requirements.

• Anthropic's finance agent templates use one markdown system prompt per agent that drives both interactive and headless modes, with no runtime prompt assembly
• Each agent splits duties across subagents (one writes, one reads, one enforces rules) to limit damage if the LLM misbehaves, mirroring bank workflows
• Data between agents is passed as regex/length-constrained JSON schemas rather than free text, blocking prompt injection from masquerading as instructions
• Guardrails like "recommend, don't decide" are prompt-based, not hard-enforced, and 3 of the 10 templates trigger EU AI Act high-risk rules on August 2, leaving roughly 20% of compliance work (risk registers, accuracy statements) still needed