Giving Agents a Mental Model of Financial Markets

Published: August 31, 2026

Since the launch of our MCP three months ago, we've been impressed by how much it's enhanced agents' equity research capabilities. Giving an agent the right set of tools can go a long way.

A tool, however, does not constitute knowledge, and at best is only a very good heuristic that produces a facsimile of what it means to understand a stock. Answers produced in this way will almost always be missing something, and spending effort to improve or add to an agent's toolkit will only compound the problem.

From tools to mental models

Instead of giving an agent a world class set of tools, we've been spending these past few weeks trying to give an agent a world class mental model of financial markets. The best way to build that model is to experience the markets yourself, carefully observing the confluence of events that occur each day. Of course, agents cannot do this (and perhaps they will never really “know” things in the way we do). But there are ways to model this, and unveiling Flexreport's knowledge graph several weeks ago was a step in the right direction.

The knowledge graph gives agents direct access to a chronological sequence of events, allowing them to “experience” a play by play for every stock in the graph. But that's not the full story. A huge part of understanding financial markets are the rules and dynamics that govern them. For example, if a company releases its latest 8-K, then it usually follows that it will also release its 10-Q, transcript, earnings deck, and so on. If an agent knows these entailments, then it will likely know what tools to use to pull this information — or temper its response if this information is not yet available.

The MCP now situates itself first

So without further ado, we are excited to announce that Flexreport Finance's MCP now operates inside of this framework: before immediately reaching for a tool to answer a question, an agent can now situate itself first by taking in the knowledge graph and the rules and relationships governing it (the ontology) to provide the most efficient and complete response.

Side by side

To see the difference this makes, we posed the same question twice — “What are the latest earnings announcements today? Any major highlights?” — once to the model out of the box (web search only), and once to an agent connected to Flexreport's MCP. Here are excerpts from each response.

Out-of-the-box LLM

Web search only — no knowledge graph

“Today's slate is thin — the last Monday before Labor Day week. One name that matters: SAIC (Science Applications International) — fiscal Q2 2027, reported before the open: Adjusted EPS of $3.01 vs. $2.31 consensus; revenue $1.88B vs. $1.76B expected.”
“Shares up ~2.9% to around $129.60, near the 52-week high.”
“Nothing noteworthy after today's close. The week loads up starting tomorrow: Dell, Palo Alto, MongoDB, GitLab and Credo after Tuesday's close; Broadcom and Snowflake Wednesday; Lululemon and Zscaler Thursday.”

Read the full chat

With Flexreport's MCP

Situated in the knowledge graph and ontology first

“Five names on the calendar today, but only one with real substance. … ANAB and APTD.L are on today's calendar but haven't posted actuals yet — ANAB consensus is -$0.41 EPS on ~$6.7M revenue.”
“Correction on the tape. I gave you +5.6% intraday — that was wrong. … Today's actual tape: prior close $128.96, opened at $133.00 (+3.1%), faded to $128.32 by 10:00, now $129.18. Essentially flat versus yesterday's close, having given back the entire gap. … So the market took the beat and sold it, which reads quite differently from the ‘+5.6% on a beat’ framing.”
“Transcript — not recorded yet. Pending, expected by roughly 4pm ET tomorrow (Sept 1). Historical arrival rate after an earnings anchor is 77%, median lag ~16 hours. Not in the cache means not captured yet, not that the call didn't happen.”

Read the full chat

Both answers got SAIC's headline numbers right. The differences are in everything around them:

The situated agent knows what doesn't exist yet. Asked for the transcript and financial release, it answered with expected arrival times and historical arrival rates — the ontology's entailments at work — instead of searching, coming up empty, and guessing.
It reads the tape, not headlines about the tape. The out-of-the-box answer relayed a stale “up ~2.9%” from a news article; the situated agent read the intraday bars, corrected its own earlier number, and caught that the stock had given back the entire opening gap — the market sold the beat.
Coverage is complete. Five names on the calendar — including LexinFintech's clear miss and two names yet to report — versus a single-stock answer.
New events reach it as they land. An investor deck at 11:20 ET and a dividend declaration at 14:42 ET entered the knowledge graph mid-conversation, timestamped, and updated the answer.