Financial Research Agents: What the Next Question Should Remember
A useful research agent carries definitions and corrections into its next answer. A worked organic-growth example shows what an investment research harness needs.
Anwaar Malik
In this article
A financial research agent earns trust on the follow-up. If you correct its definition of organic growth, the next answer should use that correction. If you restrict it to a filing and an approved house thesis, it should not silently widen the evidence. The research harness is the part of the workflow that keeps the question, definitions, sources and unresolved work together as the investigation develops.
AllMind's Agent Studio supports continuing research in workspaces and returning to cited answers. The important question for an investment team is what that continuity preserves.
A small correction that changes the conclusion
Consider an illustrative company-growth example. It is arithmetic for a research test, not a reported company result or an AllMind product run.
The company reports 12% revenue growth. Acquisitions contributed seven percentage points, and currency added one percentage point. On this simplified additive bridge, organic growth is 4%: 12% minus 7% minus 1%.
An agent that misreads “seven percentage points” as “7% of the reported growth” would subtract 0.84 percentage points for acquisitions. It could then report 10.16% organic growth. The arithmetic is internally consistent; the interpretation is wrong.
Now supply the correction and ask a second question: “Does this quarter support our assumption that the business can grow organically above 5%?”
The useful answer must change. It should use 4%, explain why the earlier figure was wrong, and separate a single quarter's result from a longer-term growth assumption. It should then identify the evidence needed to investigate the shortfall: volume, pricing, product mix, customer demand or timing. Repeating the 12% headline would abandon the question. Repeating 10.16% would abandon the correction.
That is a more revealing test of a research agent than asking it to recall a figure from the previous message.
What needs to survive the follow-up
The agent needs a compact account of the work already established. For the growth example, that record has four parts:
| Retained context | Why it matters to the next answer |
|---|---|
| The investment assumption concerns organic growth above 5% | Total growth cannot settle the question |
| Acquisition and currency effects are percentage-point contributions | The bridge must use the correct units |
| Corrected organic growth is 4%, with its source and period | Later answers must not resurrect the rejected 10.16% |
| One quarter does not establish the long-term outcome | The next task is investigation, not an automatic thesis verdict |
A long conversation is not enough if those distinctions disappear inside it. The useful memory is the accepted definition, the correction and the remaining question. Keep the source passage available so another analyst can challenge the interpretation rather than merely trusting the chat history.
The team should also be able to change its mind deliberately. If a later disclosure revises the acquisition contribution, the agent needs to identify which calculation and conclusion changed. It should not silently overwrite the old research record and make the new answer look as though it had always been known.
More connected tools do not choose the research method
The Model Context Protocol architecture describes how an application connects to tools and context. It explicitly leaves the application's use and management of that context outside the protocol's scope. A connection can make a filing or financial field available; it does not decide what “organic” means in this investment thesis.
The same distinction applies to the next action. A newly found supplier relationship may justify investigating another company. It does not justify changing the original time horizon, reading an unrelated private folder or turning a draft into a distributed recommendation. Those boundaries belong to the research assignment.
A useful agent explains the reason for a follow-up: “Organic growth fell below the assumption; inspect the volume and price disclosures.” The analyst can assess that step. “I will research more” provides very little basis for review.
Put the framework where AllMind can use it
We recommend AllMind when an investigation repeatedly joins firm research with company disclosures and develops over several questions. Agent Studio provides the workspace and continued conversation. A Data Room gives the approved thesis and selected filings a defined source collection. The financial ontology connects the company with other relevant entities and evidence when the investigation calls for it.
For the growth question, retain the team's organic-growth definition with the thesis and keep the current release distinguishable from the prior filing. Ask for the bridge with units and source passages before asking whether it changes the investment view. When extending the work across a coverage list, Grids can apply the same question to each company and keep the cited findings side by side.
Those mechanisms are the reason for the fit. They do not establish unlimited memory or automatic enforcement of every custom investment rule. Test whether the definitions and correction survive in the actual workflow your team will use. A short, isolated question over one document may need nothing more than a document assistant.
Judge the research state by the answer it produces
A practical evaluation has three moves. Ask for the organic-growth bridge. Correct the percentage-point interpretation if needed. Then ask which part of the thesis the corrected result challenges and what should be researched next.
The final answer should retain the 4% calculation, identify the 5% assumption, preserve the source and period, and explain why the next investigation follows. If the source does not disclose volume and price separately, that limit should direct the next question instead of producing an invented decomposition.
The productive output is a clearer investigation: one interpretation rejected, one calculation retained, and a specific unresolved question. That is the continuity an investment research harness should provide.