TL;DR
Anthropic published the full architecture of 10 AI agents used by Citadel, BNY, Carlyle, and Mizuho on GitHub under Apache 2.0. The valuable thing isn’t the agents themselves — they’re wired to paid finance data feeds. The valuable thing is the structure: how the work is decomposed, how skills are written, how data is separated from logic. That same template works for an SMB owner replacing the proposals, monthly close, or client-onboarding work that eats their week.
The news almost nobody is reading correctly
On May 5th, Anthropic announced ten AI agents for financial services. Citadel, BNY, Carlyle, and Mizuho are already using them.
What Anthropic published is not a tool for hedge funds. It’s an architecture. And they uploaded it complete, open, on GitHub, under Apache 2.0:
What matters: what a bank automated is what you do by hand
Look at the list of agents they released and translate it to your world:
- Equivalent at your SMB
- Your proposals for new clients that take hours to assemble
- Equivalent at your SMB
- The research you do (or don’t do) before visiting a client
- Equivalent at your SMB
- Your monthly read of financial statements
- Equivalent at your SMB
- Your monthly bank reconciliation
- Equivalent at your SMB
- Your month-end close
- Equivalent at your SMB
- Your vendor/client validation or commercial due diligence
| Agent in the repo | Equivalent at your SMB |
|---|---|
| Pitch Agent — builds commercial proposals | Your proposals for new clients that take hours to assemble |
| Meeting Prep Agent — pre-meeting briefs | The research you do (or don’t do) before visiting a client |
| Earnings Reviewer — reads reports and updates models | Your monthly read of financial statements |
| GL Reconciler — reconciles ledger accounts | Your monthly bank reconciliation |
| Month-End Closer — month-end close | Your month-end close |
| KYC Screener — validates documents from new clients | Your vendor/client validation or commercial due diligence |
It’s not an analogy. It’s the same problem, different scale. Citadel has analysts who cost USD 200,000 a year doing what you or your accounting team do by hand every month.
Anthropic just published the recipe for both of you to automate that work.
What most people will get wrong
Three predictable mistakes you’ll see on LinkedIn this week:
1. "I’ll install the agents as-is." Wrong. The agents are wired to paid connectors (FactSet, Morningstar, S&P Capital IQ) that don’t make sense for a small business. The valuable thing isn’t the agent, it’s the structure of the agent: how the work is divided, how it delegates to sub-agents, how skills are separated from data commands.
2. "This is only for technical people." Wrong. If you can read a manual, you can read an agent. To use it inside Claude Cowork or Claude Code, you install a plugin.
3. "I need Claude Enterprise." Wrong. The plugins run on any paid plan of Claude Cowork or Claude Code. Managed Agents (autonomous mode) is in public beta on the Claude Platform.
How to apply it to your company this week
Step 1 — Identify your "most expensive process"
What recurring task takes the most hours from you each month and has a clear process? Quotes, monthly reports, client onboarding, collections follow-ups, competitive analysis.
That’s your first automation candidate.
Step 2 — Copy the structure
Say you want to replicate the proposals agent — the "Pitch Agent."
Open the repo at plugins/agent-plugins/pitch-agent/ and look at how it’s organized:
- A file
agents/pitch-agent.mdwith the role and the rules - A folder
skills/with step-by-step methods - A manifest
.claude-plugin/plugin.json
Copy that structure. Replace the content with your business. That’s the mental template worth the trip.
Step 3 — Connect your data, not theirs
Where the repo connects to Bloomberg, you connect to QuickBooks, Xero, HubSpot, Google Drive, or your proposals folder. MCP is the protocol; the connectors are interchangeable.
Step 4 — Start supervised, scale to autonomous
The repo says it explicitly: every output stays in human review before acting. That’s the right rule.
Start with the agent as a copilot.
When you trust its output for 30 days, you move that task to a Managed Agent and it runs on its own.
What you’re really learning
Anthropic released this information because they understand something most AI providers don’t:
The advantage isn’t in the model — it’s in the architecture of the work.
Whoever knows how to design agents — decompose workflows, write skills, connect data, decide what gets delegated and what gets approved — is going to have a disproportionate advantage over the next three years.
That knowledge used to be locked inside consultancies and banks. Now it’s in a public repo.
If you own a small business, you have two options:
- Wait for a SaaS to appear and sell you this repackaged for USD 500+/month.
- Learn it now by copying and understanding an architecture that cost millions to develop.
The second option takes a weekend of reading and a week of implementation.
Takeaway
The important thing about Anthropic’s announcement isn’t the ten financial agents. It’s that they published the exact blueprint for how to build them. And that blueprint works just the same for a USD 50 billion fund as it does for your family business.
The question is: can I describe my work clearly and step-by-step enough for an agent to do it?
That’s an owner’s skill.
Resources
- Official announcement: anthropic.com/news/finance-agents
- Public repository: github.com/anthropics/financial-services
*Disclaimer: this article does not constitute financial, legal, accounting, or investment advice. Automating accounting or financial processes in your company should be reviewed by a qualified professional in your jurisdiction before going to production.*
Frequently asked questions
What did Anthropic publish in May 2026?
Anthropic released ten AI agents for financial services — Pitch Agent, Meeting Prep Agent, Earnings Reviewer, GL Reconciler, Month-End Closer, KYC Screener, and others — on GitHub at github.com/anthropics/financial-services under the Apache 2.0 license. Citadel, BNY, Carlyle, and Mizuho are early users.
Why does this matter for a small business?
The work the agents automate — building proposals, preparing for meetings, reading reports, reconciling accounts, closing the month, validating new-client documents — is the same work an SMB owner or accountant does by hand. The architecture is portable; only the data sources change. Citadel and your family business automate the same problem at different scales.
Do I need Claude Enterprise to use this?
No. The agent plugins run on any paid plan of Claude Cowork or Claude Code. Managed Agents (autonomous mode) is in public beta on the Claude Platform. You don’t need an enterprise contract to learn the architecture and adapt it.
Where do I start as a small business owner?
Pick your most expensive recurring process — the one that takes the most hours per month and has a clear procedure. Open the matching agent in the repo (e.g., Pitch Agent for proposals), copy its structure (agents/, skills/, manifest), replace the data connectors with your tools (HubSpot, QuickBooks, Google Drive instead of Bloomberg), and start with the agent as a supervised copilot. Move to autonomous Managed Agent only after 30 days of clean output.
Related articles
How to Decompose Any Workflow Into an AI Automation: A 7-Step Industrial Engineering Guide
A practical, industrial-engineering approach to converting manual processes into AI automations. Includes the 7-step decomposition framework plus four worked examples: web/email orders, prospect search, production intake, and transportation dispatch.
I'm Not Technical. How Do I Implement AI in My Business Before It's Too Late?
Everyone says AI is the biggest opportunity since the internet. But how does a regular business owner actually use it? Here's the honest answer — and your first step.
AI as Your Co-Founder: How We Use Artificial Intelligence to Turn Business Ideas Into Reality
You don't need a team of 50 or a million-dollar budget. With AI as your partner, you can build anything. Here's how we do it — and how you can too.