AI Agents and Agentic AI: Loops, Planning, Memory, Multi-Agent Systems and When to Use Them

import json, re, duckdb, pandas as pd, numpy as np
sales = pd.read_csv("superstore_sales.csv", parse_dates=["order_date"]); con = duckdb.connect(); con.register("sales", sales)
def sql(q): return con.execute(q).df().round(2)
TOOLS = {
    "profit_by": lambda dim, year=2025: sql(f"SELECT {dim}, ROUND(SUM(profit)/1e6, 2) AS profit_m, ROUND(100*SUM(profit)/SUM(sales), 1) AS margin FROM sales WHERE YEAR(order_date) = {int(year)} GROUP BY 1 ORDER BY profit_m"),
    "discount_effect": lambda sub_category: sql(f"SELECT CASE WHEN discount = 0 THEN '0%' WHEN discount <= 0.2 THEN '1-20%' ELSE '>20%' END AS band, COUNT(*) AS n, ROUND(100*SUM(profit)/SUM(sales), 1) AS margin FROM sales WHERE sub_category = '{sub_category}' GROUP BY 1 ORDER BY 1"),
    "trend": lambda sub_category: sql(f"SELECT YEAR(order_date) AS yr, ROUND(SUM(profit)/1e6, 2) AS profit_m FROM sales WHERE sub_category = '{sub_category}' GROUP BY 1 ORDER BY 1"),
}
class Brain:
    """Stand-in for the LLM policy: looks at the goal and the observations so far and picks the next action."""
    def next_action(self, goal, scratch):
        done = [s["action"] for s in scratch]
        if "profit_by" not in done: return {"thought": "First find where we lose money.", "action": "profit_by", "args": {"dim": "sub_category"}}
        worst = scratch[0]["observation"].iloc[0]["sub_category"]
        if "trend" not in done: return {"thought": f"{worst} is the biggest loss-maker. Is it getting worse?", "action": "trend", "args": {"sub_category": worst}}
        if "discount_effect" not in done: return {"thought": f"Check whether discounts explain the {worst} losses.", "action": "discount_effect", "args": {"sub_category": worst}}
        d = scratch[2]["observation"].set_index("band")["margin"]
        return {"thought": "I have enough evidence.", "final": f"{worst} lost ₹{abs(scratch[0]['observation'].iloc[0]['profit_m'])}M in 2025. Margin is {d.get('0%')}% without discount "
                f"but {d.get('>20%')}% when discounts exceed 20%. Recommendation: cap {worst} discounts at 20% and review pricing."}
def run_agent(goal, brain, max_steps=6):
    scratch = []
    for step in range(1, max_steps + 1):
        decision = brain.next_action(goal, scratch)
        print(f"step {step} THINK: {decision['thought']}")
        if "final" in decision: return decision["final"]
        obs = TOOLS[decision["action"]](**decision["args"])
        print(f"        ACT: {decision['action']}({decision['args']}) → {len(obs)} rows; first: {obs.iloc[0].to_dict()}")
        scratch.append({"action": decision["action"], "args": decision["args"], "observation": obs})
    return "Stopped: step limit reached without an answer."
print("\nFINAL:", run_agent("Why is our furniture business losing money and what should we do?", Brain()))

def reflect(answer, scratch_numbers):
    issues = []
    nums = set(re.findall(r"-?\d+\.\d+|\d+", answer))
    unsupported = [n for n in nums if n not in scratch_numbers and n not in {"20", "2025"}]
    if unsupported: issues.append(f"numbers not found in tool results: {unsupported}")
    if len(answer.split()) > 80: issues.append("too long")
    if "recommend" not in answer.lower(): issues.append("no recommendation")
    return issues or ["passes checks"]
draft = "Tables lost ₹0.29M in 2025 and margins fell 12.5 points; recommend capping discounts."
known = {"0.29", "2025", "-5.4", "-20.6"}
print(reflect(draft, known))