What makes a system "agentic"? The Perceive → Reason → Act loop, four architecture patterns, and where MCP fits in.
Every agent, no matter how simple or complex, runs the same loop:
A fixed sequence of steps defined by a human, with an LLM doing the reasoning inside each step. Predictable and auditable — the right default for anything regulator-facing.
The agent decides its own next step at each turn, without a fixed script. More flexible, harder to audit — used sparingly in finance, usually only for research/drafting tasks with a human review gate.
A deterministic skeleton (e.g., "always check GL, then bank statement, then classify") with autonomous reasoning inside individual steps. The most common pattern for finance agents in practice.
Several specialized agents (a Break Triage agent, a Credit Review agent, a Reporting agent) coordinated by an orchestrator that routes work to the right specialist.
| Architecture | Finance Example |
|---|---|
| Workflow | Fixed month-end close checklist: pull trial balance → reconcile → flag variances > threshold → notify preparer, in that exact order every cycle |
| Autonomous | An agent researching the current regulatory capital treatment of a new product type before a human analyst reviews the draft summary |
| Hybrid | A reconciliation agent that always follows the same 4-step check sequence, but reasons freely within each step about how to classify an unusual break |
| Multi-Agent | An orchestrator that routes an incoming request to either the AML Screening Agent, the Expense Policy Agent, or the Close Reporting Agent based on the topic |
The Model Context Protocol (MCP) is how an agent connects to real systems — your core banking platform, a document repository, a market data API — instead of just talking about them. Module 4 and the MCP deep-dive cover this in depth; for now, the important idea is: an agent's usefulness is capped by what it can actually reach, and MCP is the standard way to extend that reach safely.