From Request to Reviewable Work Product
A named agent in anthropics/financial-services owns a workflow, not an unrestricted claim on the whole financial process. The useful question is what happens between a bounded request and a file or note that a professional can inspect. This chapter follows that lifecycle: scope the ask, gather evidence, invoke the relevant methods, assemble an artifact, check it, and stop at a review boundary. The same sequence can run in a live session or be initiated by a headless event, but the transitions and ownership must remain explicit.
End to end means owning the chain between input and artifact
The named-agent prompts define a workflow in terms of an input shape and an output set. Pitch Agent starts with a target company and a strategic situation. It scopes the relevant trading comps and precedents, pulls data, builds valuation analyses, produces a football-field summary, and populates a deck. Its end-to-end responsibility is not “answer questions about a company.” It is to assemble a coherent first-draft pitch whose numbers trace back to a workbook and whose deck can undergo quality control.
Earnings Reviewer begins differently. The input is a ticker and reporting period. The agent pulls the earnings print, filings, consensus data, and the full call transcript; extracts guidance and tone; updates the coverage model; runs model QC; and drafts a note with a variance table. Market Researcher starts with a sector or theme and a one-line angle, identifies the universe boundary, builds an overview and competitive landscape, spreads the peer set, creates an ideas shortlist, and assembles a research note. Each workflow has its own sequence because the evidence and decision context differ.
The common feature is that the named agent owns the transitions between methods. It decides when a sector overview should precede a comps spread, when a model audit should follow an update, or when a reconciliation critic should re-check a break before a resolver formats it. A skill can tell the agent how to perform one operation. The named workflow determines whether that operation is appropriate now, what must be available first, and what artifact should carry the result forward.
This is also where guardrails become operational. Pitch Agent has no email or messaging tools, so its chain ends at a staged deck. Earnings Reviewer never publishes. GL Reconciler produces an exception report rather than posting an adjustment. KYC Screener recommends a risk disposition rather than deciding it. The agent owns the preparation path up to a defined state; it does not own every transition that could follow.
Those states make rework more intelligible. If a banker changes the strategic situation, the pitch should not be treated as the same run with a different label: the peer set, valuation methods, and narrative may need to be reconsidered. If a thesis changes after an earnings call, the model update and note may be correct for the old thesis but no longer sufficient for the new one. A well-bounded workflow can therefore accept a revision without pretending that every earlier artifact remains valid. The repository's handoff and review language points toward that behavior, even though the firm's own system must decide how versions and superseded files are stored.
Scope and source retrieval are lifecycle states, not preliminaries
A financial workflow can fail before its central calculation begins.