All posts
· 11 min read ·

Anthropic packaged 10 finance agents from Skills, MCP, and subagents. The stack just got productized.

The May 6 release of 10 prebuilt finance agents for banks and insurers is the first major productization of the three-layer agent stack. Each agent is described as a composition of Skills, MCP connectors, and delegated Claude subagents, with the data integration layer pre-wired by Anthropic. Early adopters include BNY, Citadel, FIS, Carlyle, Mizuho, and Travelers.

On May 6, 2026, Anthropic shipped ten prebuilt financial-services agents targeting banks, insurers, and capital-markets firms. The list reads like an internal job-board ad for an analyst pool: pitch builder, meeting preparer, earnings reviewer, market researcher, model builder, general-ledger reconciler, month-end closer, valuation reviewer, statement auditor, and KYC screener. Anthropic claims they can be deployed in days rather than months. Early adopters include BNY, Citadel, FIS, Carlyle, Mizuho, and Travelers.

What's structurally interesting isn't that Anthropic shipped finance agents. It's that each agent is described as a composition of the three layers we walked through last month: Skills for procedural knowledge, MCP servers and connectors for external access, and subagents for delegated reasoning. The agent stack just got productized.

What's in the catalog

Five agents target research and client-coverage workflows: a pitch builder that drafts presentation materials, a meeting preparer that compiles briefing packs, an earnings reviewer that processes quarterly filings, a market researcher for sector and competitive analysis, and a model builder for valuation and forecasting work.

Five agents handle operations and controls: a general-ledger reconciler, a month-end closer, a valuation reviewer, a statement auditor, and a KYC screener. The operations side maps cleanly to the work financial institutions outsource to large analyst pools today; the research side maps to what investment banks staff junior teams to do.

The product framing (Claude as Wall Street's analyst pool) is the headline. The architecture under the framing is the story.

How each agent is actually built

Anatomy of one prebuilt finance agent: pitch builder example A hierarchical decomposition diagram showing the three-layer composition of one example prebuilt finance agent. The root node is the pitch builder agent. Below it sit three child layers: a Skills layer containing procedural instructions for pitchbook drafting, an MCP connectors layer with named data integrations (Dun and Bradstreet, Fiscal AI, Financial Modeling Prep, Moody's), and a subagents layer with delegated Claude models for comparables selection and methodology checks. Each layer is annotated with what it contributes to the agent's behavior. EXAMPLE: PITCH BUILDER · ANATOMY OF ONE PACKAGED AGENT pitch builder packaged finance agent Skills PROCEDURAL pitchbook template section ordering tone + format rules

SKILL.md files loaded on trigger

Connectors MCP · EXTERNAL DATA Dun & Bradstreet Fiscal AI · FMP Moody's MCP app

MCP servers for company + market data

Subagents DELEGATED CLAUDE comparables select methodology check narrative draft

specialist Claude models called by main agent

The pitch builder example is illustrative but the pattern repeats across the catalog. Each prebuilt agent is described by Anthropic as having three composable elements: Skills, MCP connectors, and subagents — "additional Claude models that are called upon by the main agent for specific sub-tasks such as comparables selection or methodology checks." That is the same three-layer model the broader Anthropic ecosystem already shipped as a developer-facing primitive. Anthropic just made it a productized customer-facing one.

Why this is a productization moment, not a vertical launch

Most coverage frames Anthropic's announcement as a finance industry play. That framing isn't wrong (the named adopters are real financial institutions doing real analyst-pool work) but it misses what's structurally new. Two things matter beyond the industry choice:

  • The agents are composable products. Each one is a specific combination of Skills, MCP connectors, and subagents wired together by Anthropic. The wiring is not bespoke per-customer; it is a productized template that can be deployed in days. The agent web's developer-facing primitives have crossed into packaged products.
  • The data integrations are part of the package. Anthropic's list of named data partners (Dun & Bradstreet, Fiscal AI, Financial Modeling Prep, Guidepoint, IBISWorld, SS&C IntraLinks, Third Bridge, Verisk, plus Moody's via its MCP app covering 600M+ companies) is the connectors layer shipping pre-integrated. A bank that wants pitch builder doesn't have to assemble the data layer themselves; the package includes it.

The first observation is the one that generalizes. Today it's finance agents. The same composition pattern works for legal (statute researcher, contract reviewer, redline auditor), healthcare (records summarizer, claims checker, prior-auth filler), customer support (ticket classifier, escalation router, knowledge-base updater), and any other vertical with structured workflows and external data sources. The composition primitive is general; the verticals are where it gets named.

The maturity inflection

The agent web has been a developer-facing infrastructure story for most of 2025 and 2026: a developer assembles Skills, picks MCP servers, configures orchestration, and ships an agent. The economics we walked through last week imply that assembly cost is non-trivial — observability, orchestration, and memory dominate the line items. Anthropic's product is a bet that this assembly cost can be amortized across customers if the vertical workflow is standardized enough.

That bet has direct precedent in cloud infrastructure. AWS's first years were raw compute and storage; the productization moment came when industry-specific managed services (Amazon Connect for contact centers, AWS for Health, AWS for Financial Services) packaged the assembly. The agent ecosystem is at the same inflection: the productized verticals will spin out from the developer-facing primitives, and the open question is how many of them ship as single-vendor packages versus how many emerge as cross-vendor patterns.

What this means for the registry layer

A registry layer that catalogs only MCP servers and Skills (or only A2A agents) is one layer below where packaged-agent buyers are now making decisions. The decision an institution makes about which KYC-screener agent to deploy involves the underlying components but also the vendor's wiring of them: which Skills got bundled, which data partners are connected, which subagents handle which checks. That bundle is the comparison object.

The Agenstry-relevant observation, then, is that a registry needs to publish at both the component layer (raw MCP servers, raw Skills, raw A2A agents) and the package layer (this specific assembled agent for this specific use case). A consumer reading a card today sees the component. A consumer comparing the KYC-screener offerings from Anthropic, OpenAI, Google, and the smaller fintech-AI startups wants to compare assembled packages. Two different registry schemas, both useful, both currently underserved by the public indexes.

What we're watching

Three things, observable within the next two quarters:

  1. Whether OpenAI and Google ship competing packaged-agent catalogs for the same verticals. The pattern Anthropic established maps cleanly onto every frontier lab's existing primitives. OpenAI Agents SDK, Google Vertex AI Agent Builder, and AWS Bedrock AgentCore all have the component layers. The first competing finance-agent catalog will turn this from a single-vendor offering into a market.
  2. Whether the AAIF MCP Registry or a parallel index begins listing packaged agents alongside individual MCP servers. A consumer wanting to compare a KYC screener across vendors needs the bundle described. Today, that comparison happens via vendor docs and partner blog posts. A registry surface for packages is the obvious next layer.
  3. Whether the named adopters publish their own probe data on how the packaged agents perform. BNY, Citadel, Carlyle, and the others have the operational data to validate Anthropic's "days rather than months" claim. The first one to publish independent measurement will be the field's most useful external check on the productization story.

The headline read of Anthropic's announcement is "AI agents for finance." The structural read is "the three-layer agent stack just crossed from developer infrastructure into productized vertical packages." The first time that happens is interesting. The next twelve months of who copies the pattern, and how it shows up in adjacent verticals, will tell the field whether this is the beginning of the agent-as-a-service category or the largest single-vendor launch of one.

Sources

← Back to blog Agenstry