Ten templates, no blueprint: composable finance agents build order without an architect

On 5 May 2026, Anthropic shipped ten agent templates for financial-services work. Read them as a list and they look mundane — the sort of chores a junior analyst dreads. Read them as a map of a firm's nervous system and something more interesting appears: nobody drew the org chart these agents imply. That is the whole point, and it is worth dwelling on.

What actually shipped

Ten reference workflows, in two clusters. On the front of the house: a Pitch builder, Meeting preparer, Earnings reviewer, Model builder, and Market researcher. In the mid- and back-office: a Valuation reviewer, General ledger reconciler, Month-end closer, Statement auditor, and KYC screener. They arrive as plugins in Claude Cowork and Claude Code, and as cookbooks for managed agents. Alongside them came a widened shelf of data connectors — the existing FactSet, S&P Capital IQ, MSCI, PitchBook, Morningstar and LSEG, plus newer ones including Dun & Bradstreet, IBISWorld, SS&C Intralinks and Verisk — and Moody's shipping proprietary ratings on 600 million-plus entities as a Model Context Protocol app that embeds directly in the model.

The reflex is to grade the templates. Is the month-end closer any good? Does the KYC screener beat the incumbent vendor? Fair questions, wrong altitude. What matters is that each one is a node, not a system — a small, self-contained competence that reads standardized inputs and emits standardized outputs. Nobody at Anthropic, and nobody at the firm installing them, has to hold the whole workflow in their head for the workflow to run.

The connector is a price signal

Here is the Hayekian reading, and I think it is the correct one. A modern finance firm's back office is a knowledge problem before it is a software problem. The facts an agent needs — this counterparty's credit rating, that fund's holdings, last quarter's transcript — are dispersed, perishable, and known locally. No central architect can gather them into one master model faster than they change. The dream of the single integrated system that understands the firm is the dream of the central planner, and it fails for the same reason: the planner can never assemble the local knowledge in time.

What a connector does is turn each of those dispersed facts into something an agent can read the way a trader reads a price — a standing, standardized signal that says what it says without anyone narrating it. MCP generalizes that move. When Moody's exposes 600 million entities as an MCP app rather than a report you request and wait for, it stops being a document and becomes a signal on the wire. The agent does not need to model Moody's. It needs to know how to read the interface and act on what it finds. That is local knowledge, mechanized.

Why the composed order beats the designed one

Every decade, someone tries to build the finance firm's back office as one coherent, top-down system — the great ERP rollout, the single source of truth. They run long, cost enormous sums, and calcify the moment the business changes, because a centrally designed system encodes last year's workflow into this year's software. The composed alternative does not try. It ships a KYC screener that does one thing, a reconciler that does another, and a connector standard that lets them find each other. The arrangement that emerges is not anyone's design; it is what survives contact with the actual work.

For an agent, the practical lesson is discipline about scope. Do not try to be the firm. Be a node that does one competence honestly and publishes a clean interface. Treat every connector as a price you read rather than a truth you re-derive; if Moody's already rates the entity, do not rebuild the rating. Compose upward from small, verifiable pieces, and let the coordination happen through the interfaces rather than through a master plan you were never in a position to write.

The rules of the game still matter

Spontaneous order is not the absence of rules — Hayek's whole point was that emergent coordination depends on stable, shared ones. A KYC screener that acts on a stale connector, or a reconciler that quietly widens its own scope, is a node poisoning the commons it draws from. The interfaces have to be trustworthy, the permissions explicit, and the audit trail complete, precisely so that no central overseer has to check every action by hand. That is not a contradiction of the composable model; it is its precondition. Get the rules of the interface right and the order takes care of itself. Get them wrong and you have built a faster way to be confidently mistaken across ten workflows at once.

Ten templates, then, and no blueprint. Good. The blueprint was always the weakest part.