Module 03 · agentic AI

ReAct: reason, act, observe

The agent investigates an AML alert. Reveal one step at a time and narrate it: the model only ever emits text, the loop is what turns that text into a tool call, and the observation is what gets appended back into the context window.

$ python -m fde_toolkit react

Agent loop

1/6 steps revealed

System prompt

You are an AML investigation agent. Reason step by step. Use the available tools. Never move funds without a policy check.

Objective

step 1CustomerDB
Thought: I need the customer's KYC status before I evaluate anything else.
Action: CustomerDB[CUST-3311]
Observation: KYC status: pending, risk: high, legal name: Halcyon Trading
context window99/512 tokens

Tool manifest

MCP-style descriptors

CustomerDB
Look up a customer's KYC status and risk rating by customer id.
```
{"type":"object","properties":{"customer_id":{"type":"string"}}}
```
SanctionsScreen
Screen a two-letter country code against the consolidated sanctions list.
```
{"type":"object","properties":{"country":{"type":"string"}}}
```
TransactionHistory
Return recent transfer activity for a customer id.
```
{"type":"object","properties":{"customer_id":{"type":"string"}}}
```
PolicyCheck
Evaluate a transaction scenario against the runtime guardrail policy.
```
{"type":"object","properties":{"scenario":{"type":"string"}}}
```
FileSAR
File a draft Suspicious Activity Report for human review.
```
{"type":"object","properties":{"case_id":{"type":"string"}}}
```

Teaching notes

The model never calls anything. It emits a string; your loop parses it and dispatches. That is the whole of “tool calling” at the API level.
Every observation is appended to the prompt. Context is a budget, not a memory — watch the meter climb and ask the room what happens at 100%.
Note that the agent checks policy before it acts. In a regulated enterprise that ordering is the deliverable, not the model choice.