# 2026-05 — Claude for Finance and the 10 agent templates ## Why this matters On **2026-05-05** — one day before *Code w/ Claude 2026* — Anthropic shipped **Claude for Finance**: 10 ready-to-run AI agent templates for financial services, plus a finance-tuned model release (Claude Opus 4.7), plus Microsoft 365 add-ins, plus an enterprise joint-venture announcement totalling $1.5B in committed capital. For this notes set, Finance matters for the same architectural reason that [Claude Design](./2026-04-claude-design.md) matters: it's a **vertical surface built on the same agentic stack as Claude Code**, with the structure made textually explicit. [Anthropic's own announcement](https://www.anthropic.com/news/finance-agents) defines each agent template as a *reference architecture* packaging three things: > 1. **Skills** — instructions and domain knowledge for the task > 2. **Connectors** — governed access to the data the task runs on > 3. **Subagents** — additional Claude models called by the main > agent for specific sub-tasks (e.g. comparables selection, > methodology checks) That's exactly the agentic-engineering vocabulary the rest of this repo runs on: skills + connectors (the finance-vertical equivalent of MCP servers) + subagents (managed-agents-style dispatch). **The finance launch is the cleanest worked example of "agentic surface = skills + connectors + subagents" composing into a production-grade vertical product.** ## The ten agents | Agent | What it does | |---|---| | **Pitch builder** | Generates pitchbooks (the Wall Street deliverable) | | **Meeting preparer** | Prepares for client meetings: collects relevant data, drafts talking points | | **Earnings reviewer** | Reads quarterly earnings reports and surfaces what matters | | **Model builder** | Builds financial models (DCF, etc.) | | **Market researcher** | Synthesises market intelligence | | **Valuation reviewer** | Reviews valuation work | | **General ledger reconciler** | GL reconciliation (back-office accounting) | | **Month-end closer** | Month-end close work | | **Statement auditor** | Reviews financial statements | | **KYC screener** | Know-Your-Customer screening | Spread across both **front-office** (pitch builder, meeting preparer, market researcher) and **back-office** (GL reconciler, month-end closer, statement auditor, KYC screener) — Anthropic explicitly targeted both sides of the finance workforce, not just the high-billable analyst-shaped roles. What's worth noting from the [Tort Mario writeup](https://medium.com/@tort_mario/anthropics-new-ai-agents-for-finance-what-they-do-why-they-matter-and-how-they-actually-work-5379502317fd): these aren't *demo agents*. They ship as templates customers can fork, customise to their firm's data + conventions, and run as production tooling. The market shape is **template-and-customise**, not **out-of-the-box-product**. ## The Opus 4.7 finance benchmark Anthropic released **Claude Opus 4.7** alongside the finance launch and pinned its identity to a benchmark number: **64.37% on Vals AI's Finance Agent benchmark**, leading the industry. Two useful framings of that number: - **It's a finance-specific eval, not general.** Vals AI's Finance Agent benchmark is what financial-services buyers care about; it's the right number for *this* vertical, not necessarily a sign that Opus 4.7 is broadly best-in-class. (Though for Claude Code users: Opus 4.7 is the same model backing CC at the Opus tier.) - **A 64% score on a vertical benchmark is not a "solved" number.** The agent gets ~2 out of 3 tasks right. The finance launch is positioned as *accelerator for analysts*, not *replacement for analysts* — the numbers force that framing whether the marketing copy emphasises it or not. ## Microsoft 365 integration The other piece that pulls Claude further out of "chat app" territory: **Claude now works across Microsoft Excel, PowerPoint, Word, and Outlook through Microsoft 365 add-ins.** Once installed, **context moves between applications automatically** — work that starts in a spreadsheet continues in a deck without the analyst re-explaining. The cross-app context propagation is the architectural piece. It matches the same continuity-of-context shape we saw in the [Claude Design entry](./2026-04-claude-design.md) (design system read once, applied to every project) and the [CLAUDE.md entry](./2026-claude-md-and-plan-mode.md) (project context loaded once per session). **Persistent context across surfaces is the 2026 Anthropic motif** — chat, code, design, and now finance all share the pattern. ## The $1.5B enterprise joint venture Per [Fortune's coverage](https://fortune.com/2026/05/05/anthropic-wall-street-financial-services-agents-jamie-dimon/): - **$300M** each from Anthropic, Blackstone, and Hellman & Friedman - **$150M** from Goldman Sachs - Apollo Global Management, General Atlantic, Leonard Green, GIC, Sequoia Capital also participated - Total: **~$1.5B** for enterprise services This is the *bet under the bet*. The finance agent templates are a product launch; the $1.5B JV is the **distribution commitment** backing it. Goldman, Blackstone, and H&F aren't just buying — they have skin in selling enterprise-grade Claude integrations into their portfolio companies. For someone working with Claude Code, this matters indirectly: the platform's commercial trajectory determines the platform's investment trajectory. Anthropic that's $1.5B-JV-shaped invests differently than Anthropic that's only-API-revenue-shaped. ## Connection to Claude Code The finance vertical pulls into this Claude-Code-adjacent notes set because *the agent templates are constructed from the same primitives Claude Code exposes*. If you: - Read Anthropic's finance-agent reference architectures - Compared them to your own Claude Code workflows - Asked "what would my project look like packaged this way?" …you'd find your workflow already has the same pieces: skills (reusable instructions for recurring tasks), connectors (MCP servers / `--add-dir` / file references), subagents (Task tool dispatches for context-isolation). The finance agents are *your workflow, formalised as deliverables.* That's the thing worth noticing. ## Try this 1. **Read [Anthropic's finance-agents announcement](https://www.anthropic.com/news/finance-agents) for the architecture diagram.** The skills/connectors/subagents decomposition is the cleanest single statement of the agentic- surface shape Anthropic is building toward across all verticals. 2. **Pick one of the 10 agent templates.** Try to write the same thing as a Claude Code skill + plugin in your own project. You won't ship a production-grade KYC screener this way (no regulated-data connectors), but you'll feel the gap between "skill that helps a task" and "agent that runs a workflow." 3. **Read the [Vals AI Finance Agent benchmark](https://vals.ai/)** if you care about evaluation discipline. Vertical-specific evals are how this generation of agentic products will get scored; understanding the eval shape is part of agentic-engineering literacy. ## Sources - *Agents for financial services* — Anthropic (2026-05-05 primary announcement, with the skills/connectors/subagents architecture reference): - *Anthropic deepens push into Wall Street with new AI agents, full Microsoft 365 integration, Moody's data partnership* — Fortune (2026-05-05): - *Anthropic unleashes finance agents for Claude* — The Register: - *Claude for Finance: Anthropic's 10 AI Agents Hit Wall Street* — Pasquale Pillitteri: - *Anthropic's New AI Agents for Finance: What They Do, Why They Matter, and How They Actually Work* — Tort Mario, Medium (deepest single architecture writeup): - *Anthropic Launches Ten Finance Agent Templates for Claude* — Let's Data Science: - *Anthropic's Explosive Start to 2026: Everything Claude Has Launched* — Fazal, Medium (the broader release-cadence context):