The headline most people took away was automation: work that used to take three analysts and three days now takes one analyst and an afternoon.
That is the obvious story. It is not the important one.
The important story is what these agents can do on their own — and what they still cannot.
What Anthropic Actually Released
The release is best understood as a set of templates, not a finished product. Each agent is a reference architecture that bundles three things: skills, which encode the instructions and domain knowledge for a task; connectors, which give the agent governed access to the data it runs on; and subagents, smaller models called in for specific sub-tasks like comparables selection or methodology checks.
The ten cover familiar ground in capital markets and finance operations — pitch builders, meeting preparers, earnings reviewers, model builders, and market researchers on the coverage side; valuation reviewers, ledger reconcilers, month-end closers, statement auditors, and KYC screeners on the operations side. Each ships as a plugin inside Claude's desktop tools and as a cookbook for autonomous, managed agents. Claude now also works directly inside Excel, PowerPoint, Word, and Outlook, carrying context between them so a model that starts in a spreadsheet can finish as a deck without anyone re-explaining the work.
This is a serious release. It is also a clear signal about where the competition in applied AI is moving.
The Generic Workflow Is Becoming Infrastructure
For two years, the assumed advantage in applied AI was access to the best model. That assumption is aging quickly.
When one firm publishes ten working agents that any institution can fork, adapt, and deploy in days, the workflow itself stops being a differentiator. The pitchbook agent is not a moat. Neither is the reconciliation agent. They are becoming infrastructure — shared, standardized, available to everyone on the same terms.
This is the predictable pattern. The parts of knowledge work that look the same across every firm are exactly the parts that get automated first. The advantage does not disappear. It moves — toward whatever does not generalize.
Two things do not generalize: the data an agent has never seen, and the judgment no agent can hold.
The Data an Agent Has Never Seen
Anthropic's own framing is blunt: an agent is only as useful as the data and context it can reach. The release leans on exactly this. Alongside the agents, Anthropic expanded its connector ecosystem — FactSet, S&P Capital IQ, MSCI, PitchBook, Morningstar, a new Moody's app, and more. The agents are powerful precisely because they are wired into decades of structured financial data.
Now look at what is in that ecosystem. It is the established financial stack: market data, fundamentals, filings, ratings, credit graphs.
None of it carries nature.
There is no connector that tells a valuation agent that a holding depends on a watershed under stress. Nothing in a pitch agent's feeds flags that a portfolio company's supply chain runs through a region undergoing permanent forest conversion. An earnings agent can parse a transcript line by line and never surface that an asset sits on the wrong side of a tightening deforestation rule.
This is not a flaw in the agents. It reflects what exists to connect them to. As Resōno has argued, biodiversity has been largely invisible inside capital allocation models — not because it is immaterial, but because the data was never structured in a way finance could use. A more capable agent does not close that gap. It produces a more fluent version of it.
An agent can generate a pitchbook because the pattern lives everywhere in its training data. It cannot generate a portfolio's biodiversity exposure, because that intelligence does not yet exist in a form it can read. You cannot synthesize data that was never assembled.
The Judgment No Agent Can Hold
The second thing that does not generalize is harder to name, and matters more.
A fiduciary's judgment about a specific client's mandate is not a pattern to be predicted. It is a relationship. When an advisor decides that one client's foundation cares most about freshwater systems while another's portfolio is built around community outcomes, the advisor is not retrieving an answer. They are exercising judgment — and standing behind it.
No agent does that. An agent can draft, retrieve, and reconcile. It cannot be accountable.
This is why the durable layer in impact intelligence is not data alone. It is data plus judgment — the scarce, structured intelligence underneath, and the human interpretation on top that an advisor can configure, defend, and own. Strip either one out and an agent eventually absorbs what is left. Together, they are the part of the work that does not commoditize.
Where Resōno Fits
This is the layer Resōno is built to be — and it is deliberately two layers, not one.
The substrate
An immutable, science-anchored nature score, traceable to its sources and reproducible on demand. This is the part no agent can generate, because it depends on the dataset and methodology Resōno has assembled rather than on patterns a model has already seen.
The judgment layer
A configurable view where an advisor expresses each client's specific mandate — weighting what matters, incorporating the broader impact dimensions a client actually cares about, and standing behind the result. The science stays fixed. The interpretation is the advisor's, and it is fully auditable.
That structure is not a report an analyst files away. It is an intelligence layer that sits on top of an advisor's existing infrastructure and is co-built with them — human-first, with the technology as the enabler rather than the point. As agents take over the generic workflow, this is precisely where an advisor's differentiation moves: to the nature intelligence their tools can't see, and the judgment their tools can't hold.
The agent runs the workflow. The substrate and the judgment decide whether the workflow is worth anything.
Looking Ahead
The release of ten finance agents is not the end of a race over models. It is the start of one over everything the models still depend on.
The firms that treated environmental signal as a reputational overlay are about to find that their most capable new tools are blind to it. The firms that treated nature as financial infrastructure — and kept a human's judgment in the loop — will find their agents can reason about risks their competitors cannot even render.
Capital is moving from narrative to evidence, and now from evidence to agents that act on it. The constraint at every step is the same: an agent is only as good as the data beneath it and the judgment above it.
That is the layer Resōno is building.