How an agentic platform works in wealth: The Mind, The Agents, The Flow

"Agentic platform" is a term that appears with growing frequency in technology presentations for the financial sector. Like every term that gains adoption speed, it risks losing precision along the way. This essay is the opposite of a pitch — it is a technical and functional description of how an agentic platform for wealth management works, written for those who will evaluate the architecture, not for those who need to be convinced that the category exists.

The structure has 3 layers: The Mind (persistent memory graph), The Agents (coordinated autonomous workforce) and The Flow (end-to-end commercial cycle). At the end, we walk through a real example — a WhatsApp message that arrives at 14:23 and what happens in the next 60 seconds.

The Mind — persistent memory graph

The starting point for any effective agentic platform in wealth management is solving a fundamental problem: agents need context to act well. Without context, an agent is as useful as a generic model — and generic models do not know who your client is.

The Mind is a persistent knowledge graph — a structured representation of entities (clients, family members, holdings, assets, responsible bankers, events) and the relationships between them. It is not a flat relational database. It is a semantic network that can be traversed: from the client to the portfolio, from the portfolio to the assets, from the assets to the market events that affect them, from the events to the recorded interactions, from the interactions to the captured emotional and strategic context.

What the graph contains

• 01 Client profile — registration data, suitability classified with granularity (not merely "conservative" but what that means for this client: which assets they accept, which they reject, the horizon, the emotional context of risk).
• 02 Holdings and positions — full wealth structure: assets under custody, equity stakes, holdings, offshore, real estate where relevant, with updates via custody feeds.
• 03 Interaction history — not merely "call — OK" but the structured content of each relevant conversation: topics discussed, concerns expressed, commitments made, tone and emotional context.
• 04 Family and wealth network — spouse, children, heirs, partners, relationships that have implications for wealth management and for suitability and compliance decisions.
• 05 Connected market events — news and events mapped to affected positions, with relevance calculated per client. Not the 500 news items of the day — the 3 that matter to this specific client.

The graph is fed from multiple sources: custody feeds, processed WhatsApp transcripts, banker meeting notes, filtered news feeds and scheduled enrichment by specialized agents. Each new piece of data is inserted with timestamp, source and context — creating an auditable timeline of everything the system knows about each client.

"The Mind is not a database with search. It is what a banker with 15 years at the house knows — but available to any agent, instantly, with no degradation from turnover."

The Agents — coordinated workforce

The second pillar is the agentic workforce — a set of specialized agents that operate over the memory graph, coordinated by a central orchestrator. Each agent has a well-defined scope of responsibility, delimited permissions and a specific mode of operation (active or reactive, scheduled or event-driven).

Concierge — the entry point

The Concierge is the intake agent — the point of contact between the external world and the system. Every inbound message passes through the Concierge: client WhatsApp, email, banker request. The Concierge identifies the intent, classifies urgency, resolves what it can resolve within its permissions and forwards to specialized agents what requires additional processing.

Oghma — the orchestrator

Oghma is the central orchestrator — the agent that receives complex tasks and decomposes them into subtasks distributed among the specialized agents. When a meeting is scheduled, Oghma does not generate the briefing alone — it instructs the portfolio agent to update positions, the suitability agent to verify the profile, the news agent to filter relevant events, the compliance agent to check open items. It receives the results and synthesizes the final briefing.

Oghma maintains the state of each task in progress, manages dependencies between subtasks and ensures that failures in individual agents do not block the final result. It is LLM-agnostic — it can orchestrate agents running on Anthropic Claude, OpenAI GPT, open-source models, or a combination.

Specialized agents

• Investments Agent — accesses positions from the graph, calculates exposure, identifies concentration, compares with benchmark and the client's investment policy. Produces structured portfolio analysis, not generic narrative.
• Suitability Agent — verifies whether a proposed communication, suggested allocation or offer is within the suitability profile classified for that specific client. Flags divergences before they reach the banker.
• Compliance Agent — applies rules configured by the institution's compliance officers across all outbound communications. Checks Anbima, CVM and internal-policy restrictions. Generates an audit trail of each decision.
• Briefing Agent — synthesizes every input from the other agents into a structured document, readable in under 60 seconds, in the format the banker recognizes and uses.
• News Agent — processes news feeds and filters by relevance to each client's portfolio. It does not send what is "important news" — it sends what is important to this client, based on the assets they hold.
• WhatsApp Analyst — processes inbound WhatsApp messages: transcribes audio, identifies intent, extracts commitments made, updates the graph with interaction content, proposes a reply for banker review.

"Each agent has delimited permissions. No agent makes decisions outside its scope. The banker remains in the approval line for every action that has external consequences."

The Flow — a real example

Theory is necessary but insufficient. We walk through a concrete example: a WhatsApp message that arrives at 14:23. The client is Mr. Monteiro, patriarch of a family with wealth under management at a bank with private banking. The message reads: "I need to talk about what is happening with the fund. I am concerned."

14:23:01 — Message received

The message enters through the WhatsApp Business API integration channel. The Concierge receives it, identifies the sender as Mr. Monteiro (via number mapped in the graph), classifies the intent as "concern about performance — requires urgent briefing" and triggers Oghma at high priority.

14:23:04 — Oghma decomposes the task

Oghma reads Mr. Monteiro's context in the graph: moderate-conservative suitability profile, 42% of the portfolio in hedge funds, the last recorded interaction was 12 days ago. It decomposes into parallel subtasks: (1) Investments Agent: performance of the funds in the portfolio over the last 30 days; (2) News Agent: market events relevant to Mr. Monteiro's portfolio; (3) WhatsApp Analyst: history of previous interactions on the topic of performance.

14:23:08 — Agents in parallel

Investments Agent pulls positions from the graph and calculates: the reference fund in Mr. Monteiro's portfolio dropped 3.2% in the month, against a benchmark of −1.8%. News Agent identifies two relevant articles published in the last 48 hours about the fund's sector exposure. WhatsApp Analyst finds in the history that in March Mr. Monteiro expressed discomfort with volatility above 2% monthly and that the current banker committed to calling if the fund crossed that level.

14:23:19 — Synthesis and proposed response

Oghma receives the results from the 3 agents, triggers the Suitability Agent to confirm that a "context + suggestion to call" response is within scope, and the Compliance Agent to verify that nothing in the proposed response violates communication policies. With approval from both, the Briefing Agent synthesizes a response message for Mr. Monteiro and an internal memo for the responsible banker.

14:23:41 — Banker receives the package

The banker receives in the dashboard: (1) Mr. Monteiro's full context, updated with today's message; (2) the fund's performance analysis with market context; (3) the March commitment that was not honored — the banker should have called when volatility crossed 2%; (4) draft WhatsApp response, awaiting approval; (5) suggestion to schedule a call in the next 30 minutes.

The banker approves the message with one click, adjusts two terms, and the response is sent. The interaction is recorded in the graph. The March commitment is updated as "pending — resolved with scheduled call". All of this between 14:23 and 14:24.

Security, isolation and deployment — for the technical buyer

The questions every bank or asset-manager CIO and CTO asks at this point are predictable — and legitimate. We answer them directly:

Multi-tenant isolation

Each tenant (each bank, each division, each segment) operates with a fully isolated graph. There is no data sharing between tenants. A bank with private banking and asset management in the same institution can configure separate tenants with distinct compliance policies, with no data crossover between the divisions.

On-premise deployment

The platform runs on the institution's VPS or data center. Coolify, Kubernetes or a traditional VM — the choice is the institution's. Client data never leaves for third-party servers. The LLM the agents use can be self-hosted (open-source models) or accessed via API under enterprise contracts that guarantee data is not used for training. The institution chooses — and retains full auditability over where the data lives.

Audit trail

Every agent action is logged: which agent executed, with which inputs, with which output, at which timestamp, with which result. For CVM, Anbima and internal compliance supervision purposes, the audit trail is complete and immutable. Agent decisions that produce client communications are recorded with the full context that led to the decision.

LLM-agnostic

No agent is bound to a specific language-model provider. The institution can use Anthropic Claude for the orchestrator and a self-hosted open-source model for the compliance agent — or any other combination. This is especially relevant for institutions with regulatory restrictions on the use of specific cloud providers, or with a data-sovereignty roadmap in the medium term.

"It is not magic. It is architecture. And the architecture runs on your VPS, with your data, under your policies."