Intro.
#What Anthropic Released
In May 2026, Anthropic open-sourced the `anthropics/financial-services` repository — ten Claude agent templates purpose-built for financial-services workflows.
| Agent | What It Does |
|---|
| Pitch Agent | Builds pitch books and IR materials |
| Market Researcher | Conducts market and competitive research |
| Earnings Reviewer | Analyzes earnings and financial statements |
| GL Reconciler | Reconciles the general ledger |
| KYC Screener | Verifies customer identity |
| Valuation Agent | Values companies |
| Credit Memo Agent | Drafts credit memos |
| Fund Accounting Agent | Handles fund accounting |
| Month-End Agent | Automates month-end close |
| Financial Audit Agent | Audits financial statements |
The architecture isn't simple. It's a hierarchical structure where an orchestrator agent coordinates subordinate worker agents, complete with a steering-events mechanism for injecting user instructions in real time mid-review, per-agent permission isolation (read-only vs. write access), and external data access through MCP connectors.
02
#Why "AI Is Replacing Investment Banking" Is Only Half True
What these agents actually replace is precisely the repetitive data collection, organization, and formatting work junior bankers used to grind through. Deciding how to interpret the data, and which direction an investment decision should go, still belongs to human judgment.
Here's the more important fact. By releasing this repository, Anthropic has officially validated that domain-specialized, multi-agent architecture works in real financial workflows. A structure that works in investment banking works in startup review, too.
TIP
OpenSeed's 15-reviewer structure — 7 core reviewers, 8 specialists, and an IC chair — is fundamentally the same architectural pattern as Anthropic's financial-services repo. Only the domain is different.
03
#What OpenSeed Takes Away From This
Analyzing Anthropic's repo, we identified three patterns OpenSeed can adopt going forward.
| Pattern | How the IB Repo Implements It | How OpenSeed Applies It |
|---|
| Agent confidence scoring | Sub-agent judgments are tagged with confidence labels | The Chief reviewer dynamically down-weights agents working from thin data when computing the final score |
| Steering events | User instructions are injected in real time mid-review | Follow-up instructions like "focus on regulatory risk" get incorporated while review is still in progress |
| Source tracing | Every claim is tagged with its source agent and underlying data | You can trace the reasoning behind any judgment back to the specific agent that made it |
All three point in the same direction: making review results more trustworthy. If an AI can't explain why it reached a judgment, you have no reason to trust the result. That interpretability is the core difference between OpenSeed and a generic AI chatbot.
Summary.
#When Does the Real Threat Arrive?
There may come a day when Anthropic ships a "startup review agent template" the same way it shipped IB agents. That's when real competition begins.
But even if that day comes, the platform and the product remain different. What Anthropic builds is agent infrastructure. What OpenSeed builds is a review network trained on the realities of Korea's startup ecosystem — government grant scoring rubrics, how VC reviewers actually think, and the review vocabulary that spans everything from the Pre-Startup Package program for first-time founders to TIPS, Korea's deep-tech investment track. That context doesn't come out of a public repo.
CTA
OpenSeed's review structure is built on six years of hands-on startup experience and a 15-agent collaborative system. Start free during the current beta.
Start Your AI Agent Review Right Now
15 review agents analyze your business plan domain by domain.
🔒 Free during beta · your submission isn't saved
Start Free AI Feedback →