Anthropic quietly open-sourced a financial-services repository that maps out an entire skill-driven architecture for AI agents — and it’s more revealing than most whitepapers. The repo, cloned and parsed programmatically in a new tutorial, exposes a layered plugin system built around vertical-specific financial analysis, partner integrations, and managed-agent cookbooks. By walking through the directory tree, you can see exactly how Anthropic organizes agent capabilities: 44 distinct plugin directories split across agent plugins, vertical plugins, partner-built integrations, and cookbook-style managed agents.
The real meat, though, is in the SKILL.md files. These aren’t just documentation — they’re structured playbooks that get injected directly into the model’s context via the Messages API. The tutorial builds a SkillRegistry class that parses YAML frontmatter from every one of these markdown files and makes them searchable by name or description. Think of it as a RAG system where the retrieval corpus is Anthropic’s own opinionated financial workflows rather than generic documents.
What’s clever about this approach is how the SkillAgent wrapper handles tool use. Rather than relying on a single prompt, it maintains an iterative loop where the model can call Python functions for calculations — discounted cash flow models, WACC sensitivity analysis, comparable-company comps — and produce formatted Excel output or matplotlib heatmaps. The tutorial demonstrates this with a synthetic DCF valuation that generates a terminal-growth sensitivity chart and spits out a private-equity investment committee memo. One managed-agent deployment spec even gets inspected without triggering a live request, which is a nice touch for anyone nervous about accidentally spinning up production infrastructure during testing.
The architecture essentially reproduces what you’d get from Anthropic’s managed agent platform but in pure Python with the standard SDK. That means full visibility into how skills are selected and applied, plus the ability to modify or extend the registry. For financial analysts who’ve been hesitant about black-box agent platforms, this kind of inspectable pipeline might actually change the conversation.
