The paper announces the release of ten ready-to-run AI agent templates specifically designed for financial services, aiming to automate and streamline time-consuming tasks such as building pitchbooks, screening KYC files, and month-end closing processes. These agents are built upon Claude Opus 4.7, which is highlighted as state-of-the-art for financial tasks, leading the Vals AI Finance Agent benchmark with a 64.37% score.

The core methodology of these agent templates involves a structured reference architecture. Each template packages three key components:

1. Skills: These encompass the detailed instructions and specialized domain knowledge required for a particular financial task. This includes the procedural steps, analytical frameworks, and industry-specific nuances necessary for the agent to perform accurately.
2. Connectors: These provide governed, real-time access to the necessary data sources for the task. This involves integrations with a broad ecosystem of financial data providers (e.g., FactSet, S&P Capital IQ, Morningstar), research platforms, and firms' internal systems (data warehouses, CRMs). New connectors are introduced for Dun & Bradstreet, Fiscal AI, Financial Modeling Prep, Guidepoint, IBISWorld, SS&C IntraLinks, Third Bridge, and Verisk. An MCP (Managed Custom Process) app from Moody's is also introduced, embedding proprietary tools directly within Claude.
3. Subagents: These are additional Claude models called upon by the main agent to perform specific, granular sub-tasks. Examples provided include subagents for comparables selection or methodology checks, allowing for a decomposition of complex financial workflows into manageable, specialized AI tasks.

These agent templates can be deployed in two primary ways:

4. As plugins within Claude Cowork or Claude Code: In this deployment model, the agent operates alongside a human analyst, running on the analyst's desktop environment. It integrates seamlessly with Microsoft 365 applications (Excel, PowerPoint, Word, and Outlook via new add-ins). A core technical feature here is the automatic context propagation across these applications, meaning work initiated in one application (e.g., a financial model in Excel) carries its context to another (e.g., a presentation deck in PowerPoint) without requiring re-explanation. This also includes the "Dispatch" feature, allowing users to assign tasks to Claude for processing local files even when away from their desk.
5. As Claude Managed Agents: This deployment method enables autonomous operation of the templates on the Claude Platform. This is designed for larger-scale or scheduled work (e.g., spanning a book of deals or nightly processes). The "cookbooks" for Managed Agents provide firms with ready-made building blocks that would otherwise require significant engineering effort. These include technical features such as long-running sessions (to handle multi-hour tasks like deal closes), per-tool permissions for granular access control, managed credential vaults for secure access to external systems, and a full audit log in the Claude Console. The audit log offers compliance and engineering teams a comprehensive record of every tool call and decision made by the agent, ensuring auditable outcomes.

In both deployment scenarios, the human user remains firmly "in the loop," providing oversight by reviewing, iterating on, and approving Claude's generated work before it is finalized or externalized. The system emphasizes augmentation of human capabilities rather than full automation without human oversight. The overarching goal is to significantly compress the time required for complex financial tasks, enhancing efficiency and accuracy.