Skills, connectors, subagents: the new agent architecture, explained for finance leaders
Published 13 May 2026
When Anthropic shipped its finance agents on 5 May, the launch made the architecture explicit. Each agent template, the announcement said, packages three things: skills, connectors, and subagents. (Source: anthropic.com/news/finance-agents.)
The architecture is not just Anthropic’s. The same pattern is showing up in the agents from other vendors, and it will be the dominant pattern for the next phase of AI in finance. The finance leader who understands it does not need to be a technologist. They do need to know what each component is, what it changes, and why the pattern matters for decisions the function will be making this year.
This post is the leader’s explanation. Short, specific, and applicable.
Skills: the instruction set the agent follows
A skill is the instruction the model is following to do the work. In practice, it is a long, structured prompt that includes the steps the agent should take, the rules it should apply, the categories of judgment it should and should not make, and the escalation logic for the edge cases.
The skill is the part of the agent that maps to the use case. The “month-end closer” skill is different from the “KYC screener” skill is different from the “valuation reviewer” skill. The skill is also the part of the agent that can be customised for your function. The chart of accounts, the materiality thresholds, the reporting calendar, the documentation conventions. All of those live in the skill.
What that means for a finance leader is that the skill is the lever you have most direct influence over. The model is the vendor’s. The connector is the data partner’s. The subagents are the vendor’s again. The skill is configurable, and the way you configure it is the way the agent does your work.
The decisions the function has to make about the skill are surprisingly important and surprisingly under-discussed. Who edits it. Who versions it. Who reviews changes. Who tests the changes against a representative workload before the changes land in production. The function that treats the skill as the vendor’s responsibility is the function that does not understand what is sitting in front of its work.
Connectors: the governed routes to data
A connector is a structured pipe between the agent and a source of data. The agent does not see the world directly. It sees the data the connectors let it see, in the form the connectors return it, with the controls the connectors apply.
Connectors do three things, as I covered in the connector post. They authenticate access. They log every request. They return data in a known shape under a known contract.
What that means for a finance leader is that the connectors are the trust layer of the deployment. The agent that has access to your ledger via a governed connector is the agent whose work you can defend in audit. The agent that has access to your ledger because somebody pasted a CSV into a chat is the agent you have a problem with.
The connectors are also where the data partnerships sit. The Anthropic launch is heavy on connectors because the data partnerships are heavy. The list runs across tier-one financial data vendors (FactSet, S&P Capital IQ, MSCI, PitchBook, Morningstar, LSEG, Moody’s via MCP) and includes a fresh wave of MCP integrations (Dun & Bradstreet, Fiscal AI, Guidepoint, IBISWorld, SS&C Intralinks, Third Bridge, Verisk). The MCP design lets new connectors be added without the deep engineering work the previous generation of integrations required. The finance leader who watches the connector library expanding over the next twelve months is watching the surface area of what the agent can credibly do in your function expanding with it.
Subagents: the smaller models called for sub-tasks
A subagent is an additional model the main agent calls for a specific sub-task. The main agent is the orchestrator. The subagent does a narrower piece of work, returns the result, and the main agent integrates it.
The pattern is older than this generation of finance agents. It is how a lot of complex AI workflows have been built since 2024. What is new in the Anthropic templates is that the subagent architecture is documented as part of the template, not hidden inside it.
In a pitch builder, a subagent might do comparables selection. In a month-end closer, a subagent might do variance computation. In a valuation reviewer, a subagent might check the model’s methodology against a reference framework. The point is that the work is decomposed, and each piece is done by a model that is appropriately sized and prompted for the work.
The reason this matters for a finance leader is that the subagent architecture is where a lot of the performance and a lot of the cost live. A well-designed subagent layer makes the agent faster, cheaper, and more accurate. A poorly-designed one is where the budget goes to die. The vendor that documents the subagent architecture is the vendor that has thought about the deployment. The one that does not is asking you to trust the wrapper.
Why the pattern matters
The pattern matters because it makes the agent legible.
Before this pattern, the model was a black box and the deployment was a leap of faith. After this pattern, the agent is a set of named components, each with its own contract, its own audit trail, and its own evaluation criteria.
That changes a lot of conversations.
The vendor conversation changes. You are no longer buying a tool with features. You are buying an agent with three layers, and you can interrogate each layer. The vendor evaluation post is the long version.
The IT conversation changes. The IT function has a clear set of components to evaluate, control, and integrate. The skill, the connectors, the subagents, the audit trail. Each one maps to a piece of the IT estate the function already manages.
The audit conversation changes. The auditor has a set of components to test, with documented contracts and full audit trails. That is the foundation of the continuous attestation argument. The agent is legible because the architecture is legible.
The team conversation changes too. The team is not working alongside a single AI tool. The team is working with a system that has named parts. The conversation about what the team is responsible for, what the agent is responsible for, and what the data partner is responsible for becomes a conversation about specific layers, not about AI in general.
What this implies for your function this year
Three things.
Treat the skill as something the function owns. Even if the vendor wrote the first version, the function customises it, versions it, and tests changes to it. The skill is the part of the agent that does your work. Own it.
Treat the connector library as a strategic asset. The connectors you have governed access to determine what the agent can credibly do in your business. The conversation about which connectors matter most to your function should be on the agenda. The data partnerships that look like vendor noise are actually the surface area of your AI capability.
Insist that the subagent architecture be documented. Vendors that hide the subagent layer are selling you a black box. Vendors that document it are selling you a system you can govern. Choose the latter.
The building an AI-ready finance function post covers the broader readiness picture. The architecture question is the one that has come into focus since May.
Where this lands
The agent architecture is the most important change finance leaders will adjust to in the next two years. The change is not that finance teams are using AI. They have been doing that for a while. The change is that the AI is now a system with named components, each with its own contract, its own audit trail, and its own evaluation criteria.
The finance leaders who learn to talk about the components, not just the AI, are the ones who will make the better deployment decisions. The leaders who keep treating AI as a single black box will be the ones whose deployments slip, whose costs balloon, and whose audit trails are not adequate.
The pattern is named. The work is to apply it.
Maebh Collins is a Fellow Chartered Accountant (FCA, ICAEW) with Big 4 training and twenty years of operational experience as a founder and senior finance leader.