6
Agents composing the system
A supervisor plus specialist agents for coaching, app help, banking data, actions, and small talk.
Case Study
Multi-Agent financial coach and assistant
Vstorm built a multi-agent solution for Yeeld, a UK-based fintech startup. The system uses a supervisor architecture: a leading agent assigns tasks to more focused agents from the user's prompt and conversation.
The outcome
A financial companion built as agents, not as a single chatbot
Yeeld started from an idea, not a legacy stack. The product had to feel like an accountant, a spending coach, and a financial advisor in one app — connecting accounts, reading spending patterns, and coaching habits without inventing answers from the open web.
The figures describe the production supervisor graph on LangChain and LangGraph, not a lab wrapper around a generic chatbot. Only the agents needed for a given prompt run. The rest stay quiet, which keeps token cost down.
2
Coaches with opposite attitudes
One finds the bright side of a habit. The other flags red flags against demographic benchmarks.
1
Supervisor that routes the work
Splits the user request and calls only the specialists that prompt actually needs.
About the client
Yeeld is a UK-based startup building an AI-powered financial companion — comparable to an accountant, a spending coach, or a financial advisor — in one agentic app. The product connects accounts and other sources so it can read spending habits and point to areas to improve.
Yeeld was established in 2021 in the United Kingdom, with a vision to make a modern, convenient, AI-powered financial habit-builder and coaching app available to everyone.
Vstorm's impact
Vstorm's impact, the TL;DR
- Full agentic architecture delivered from a blank slate — no legacy system to work around
- Supervisor graph on LangChain and LangGraph, with specialist agents called only when the prompt needs them
- Two coaching agents with opposite attitudes, both reading demographic benchmarks
- Assistant agent backed by RAG over Yeeld's own docs, so app help is not guessed
- Financial agent with banking API access and chart generation
- Two habit games — Swipe the Bill and Quizzer — plus non-standard notifications on live behaviour
The challenge
The challenges of building healthy financial habits
According to Ramsey's State of Personal Finance report for 2025, 49% of Americans are living paycheck to paycheck and 50% worry every day about their personal finances. Forty-three percent of US adults report difficulty paying bills; 34% struggle to provide food. For 49% it is easier to get a loan than to build personal savings, and 34% say they are more likely to spend when stressed or emotional. A lack of financial literacy and of durable habits is a large part of that picture.
Yeeld wanted an agentic companion that could sit with that problem. Vstorm started from the idea alone. There was no incumbent chatbot to wrap, and no legacy stack to negotiate. That was the constraint and the opening: design the architecture, build the agents, design the ecosystem around them, and ship something that actually coaches.
How we delivered
TriStorm on a product that did not exist yet
The coach had to live inside Yeeld's app, not beside a generic chatbot. Delivery started from a blank slate and ended in a supervisor graph plus two habit games.
Choose the supervisor pattern
Workshops aligned the architecture with Yeeld's goal of a universal financial advisor. A leading agent would split the user request and call specialists — the way a lead engineer calls builders, electricians, and plumbers — instead of one model answering every turn.
- Supervisor brief
- Specialist map
- Blank-slate constraints
Proof of Value on real coaching turns
The graph landed on LangChain and LangGraph: a supervisor, two coaches with opposite attitudes, an assistant with RAG over app docs, a financial agent on banking APIs, an action agent, and a cheap conversational agent for hello and thank-you.
- Supervisor graph
- Two-attitude coaches
- RAG-backed assistant
Make the habit loop playable
Swipe the Bill and Quizzer went in so coaching is not only a chat. Notifications surface live behaviour. The app is agentic-first: data in, habits out, with the graph orchestrating each prompt.
- Swipe the Bill
- Quizzer
- Behaviour notifications
How it works
A supervisor, then only the specialists the prompt needs
The leading agent coordinates the rest. It splits the user request into tasks and hands each to a narrower specialist. Agents that are not needed for that turn do not run, so they do not spend tokens. The graph was built with LangChain and LangGraph.
User prompt Financial question or action
↓
Supervisor agent Routes tasks to specialist agents · LangChain + LangGraph
↓
Conversational Small talk · saves tokens
Supportive Highlights positives in spending
Cautionary Flags risks and red flags
Assistant App features via RAG-backed docs
Financial Reads banking data · generates charts
Action Maps requests to app capabilities
↓
Response to user
Vstorm × Yeeld — LangChain + LangGraph supervisor graph
| Agent | Stance or tool | What it does |
|---|---|---|
| Supportive | Optimistic | Cherry-picks positive habits against demographic benchmarks and shows the user where they are on track. |
| Cautionary | Cautious | The opposite coach: flags failures and red flags against the same demographic data. |
| Assistant | RAG over app docs | Helps the user manage Yeeld features without hallucinating answers from outside the product. |
| Financial | Banking APIs + charts | Reads connected account data and can return it as a chart or graph. |
| Action | App capability map | Maps a request to what the app can actually do, and says so when it cannot (for example, buying a particular stock). |
| Conversational | Token cost | Handles hello, goodbye, and thank you. Stays polite and declines topics outside its competence. |
Two coaches, opposite attitudes
Both coaching agents see aggregated data on the financial habits of people across demographics, so they can compare the user's decisions to a benchmark. The supportive agent is hard-wired to find the bright side. The cautionary agent is hard-wired to worry. Together they stop the product from being only a cheerleader or only a scold.
Gamifying spending
Two games sit beside the graph so habit-building is not only a conversation.
Swipe the Bill uses a left-or-right swipe, but the cards are bills and spending pulled from the user's account. Approve or condemn each payment, then get a report on how many felt worth it. The point is reflection, and a place to look for savings.
Quizzer (“Are you smarter than the rest”) asks whether the user thinks they spend more or less than a comparable population on food, clothing, electronics, services, and the rest — then contrasts that expectation with the generalised data.
Results
What Yeeld shipped
Vstorm delivered a working agentic workflow that orchestrates specialist agents on each user prompt, so the advice is specific to the account, the habit, and what the app can actually do. Pattern extraction shows where to look for improvements. Notifications surface ongoing behaviour instead of waiting for the next chat turn.
There is no before/after conversion figure on this engagement to plot. The shipped system is the outcome: a supervisor graph, two-attitude coaching, RAG-backed help, banking-aware charts, and two games that make the habit loop playable.
Work with us
Ready to see how agentic AI transforms financial coaching?
Meet directly with our founders and PhD AI engineers. We will walk through real implementations from 30+ agentic projects and the practical steps to integrate them into your workflows.