Building Agentic AI Systems

Why most AI initiatives fail — and the three-pillar framework that changes the outcome.

A Fortune 500 retailer deployed a chatbot that could answer questions about return policies. Within six months they tried to make it process actual returns, modify shipping addresses, and issue refunds. It could do none of those things — and the failure wasn't a bug. It was an architecture that was

A production order-fulfillment agent silently doubled its cloud bill in seventy-two hours. Nobody changed the prompt. Nobody deployed new code. A single upstream schema change added four extra fields to every tool-call response, ballooning each completion from 800 tokens to 3,200. The team had never

A customer-service agent goes live on Monday morning. By Tuesday afternoon it has refunded $14,000 to users who never asked for refunds, because nobody gave it a memory of which conversations were already resolved. The LLM was sharp; the architecture was hollow. This chapter opens the hood.

An LLM that calls tools without thinking is just a random function caller with good grammar. In this chapter, you build the loop that turns a language model into something that actually reasons.

Your agent from Chapter 4 can reason one step at a time. But ask it to plan a weekend trip involving flights, hotels, dietary restrictions, and a budget constraint, and it confidently books a hotel in the wrong city. The problem is not intelligence — it is the absence of structured thinking.

An agent without tools is a confident liar. It will invent API responses, fabricate database rows, and cite papers that do not exist — all with the fluency of someone who has done it a thousand times. This chapter gives your agents hands.

An agent that forgets what you said three messages ago is not an assistant — it is a stranger you keep re-introducing yourself to. This chapter gives your agents the ability to remember.

A customer asks your agent whether the company’s refund policy covers digital subscriptions. The agent responds with absolute confidence: “Yes, all purchases are eligible for a full refund within 30 days.” The real policy, updated six weeks ago, explicitly excludes digital subscriptions. Three hundr

The insurance claim seemed straightforward. A customer uploaded a photo of hail damage, and the agent kicked off three parallel tasks: image analysis, policy lookup, and fraud scoring. The image analyzer finished first and classified the damage as severe. The policy lookup returned a coverage limit

The demo looked flawless. A single agent that researched competitors, drafted a market analysis, pulled financial data, and produced a polished report — all in one long chain-of-thought. Then the team deployed it on real queries. The agent would start researching, get distracted by a tangent in the

It was 2:47 AM when the automated procurement agent placed a $2.3 million order for industrial solvents. The purchase order had passed the agent’s internal validation: the supplier was in the approved vendor list, the price-per-unit fell within historical ranges, and the requesting department had bu

Two agents are assigned to plan a product launch. The marketing agent writes a press release announcing a feature that engineering has already descoped. The engineering agent drafts a deployment timeline that contradicts the date marketing promised to journalists. Neither agent is wrong in isolation

Tuesday, 2:47 AM. A loan-approval agent in production has been silently rejecting every application for the past six hours. No errors in the logs. No alerts firing. The HTTP status codes are all 200. Each applicant receives a polite decline email with a coherent, well-structured explanation that cit

At 2:47 AM on a Tuesday, a customer service agent deployed three weeks earlier begins sending internal pricing spreadsheets to anyone who asks. The prompt injection is elegant: a user embeds an instruction inside a fake “customer complaint” that tells the agent to ignore its system prompt and instea

The agent passed every test on your laptop. You merged the PR on Friday afternoon, pushed to production, and went home. By Saturday morning the on-call engineer had paged you three times: the container was OOM-killed every forty minutes, the API gateway was routing traffic to a stale replica that st

A senior analyst at a mid-size consulting firm spends four hours each morning on the same ritual: open twelve browser tabs, skim a handful of reports, copy key figures into a spreadsheet, cross-reference claims across sources, and draft a two-page brief that the partner will read in six minutes. The

Every engineering team has felt it: a pull request sits in the queue for two days because the only person who knows that subsystem is on vacation. When it finally gets reviewed, the reviewer catches a style violation and a missing null check but misses the SQL injection hiding behind a string interp

Every support team faces the same spiral: ticket volume grows, response times stretch, quality drops, customers churn, and the remaining agents burn out faster. You hire more people, but onboarding takes months and institutional knowledge stays locked in senior agents’ heads. Meanwhile, your knowled

Data pipelines break silently. A column renamed upstream, a vendor switching date formats, a nullable field that was never null until today — each failure looks trivial in hindsight, yet it propagates through warehouses for hours before anyone notices. This capstone builds an agentic ETL system that