Deep Dive

Engineering AI Agents

Bounded tool-using loops with durable state, guardrails, and evals

Agents are not chat completions in a while loop. This Deep Dive teaches the engineering discipline behind production agents: bounded execution loops with explicit termination conditions, governed tool portfolios, durable memory and checkpoints, recoverable failure handling, and evaluation methods that measure real agent behavior instead of vibes.

Latest Updates 2026

See the Invisible

Interactive simulators visualise what's hidden from view.

Hands-On Labs

Step through executions tick by tick. Manipulate state.

Why, Not Just What

Understand the reasoning behind every design decision.

Quizzes & Cheatsheets

Verify your understanding and keep a quick reference handy.

Get Certified

Earn a shareable certificate to prove your deep expertise.

The AI Era Demands More

Become the Engineer Who Supervises AI

As AI generates more code, understanding what that code does becomes more valuable, not less. Someone must verify AI output, debug failures, and make architectural decisions.

Build Your Architectural Edge

From a demo that impresses to an agent you can put in production

A tool-using agent looks finished after the first successful run: it picks tools, chains steps, and reaches the goal. Then real traffic arrives, and the same agent loops without making progress, retries an irreversible action, burns its token budget on a dead end, and leaves no trace of why. This Deep Dive treats the agent as a system you engineer: you decide which choices the model makes and which stay in deterministic application control, you bound the loop with budgets and termination conditions, you classify actions by how badly they can go wrong, and you build the evaluation harness that tells you whether a change made the agent better or just different.

What You Will Be Able to Do

Design a loop that always terminates

Define goals, observations, and actions for an agent, set step, token, time, and cost budgets, and choose termination conditions and stop reasons so every run ends in a clean terminal state instead of an unproductive loop.

Make tool use safe by construction

Classify each action as read-only, reversible, or irreversible, scope least-privilege credentials per tool, place approval gates on consequential actions, and pair retries with idempotent actions and compensating actions so a failed run does not corrupt real systems.

Prove agent behavior with evals

Assert on task outcomes, constraints, and side effects, analyze trajectories and tool selection, inject tool faults and adversarial instructions, and use nondeterministic trial sets to state a success rate you can defend.

What's Covered

The execution loop and its boundaries

The cycle of goals, observations, actions, and environment feedback, how agent state persists across repeated model invocations, and the line between model decisions and deterministic application control, including when a fixed workflow beats dynamically selected actions.

Governed tools and planning

Minimal tool portfolios with preconditions and postconditions, dependency-aware ordering, task decomposition into verifiable subgoals, and replanning when observations contradict the plan, with search breadth and execution depth traded against budget.

Memory that earns its keep

Working context versus durable memory, episodic records, semantic facts, and procedural state, write criteria with provenance, retrieval scoped to the current goal, and the operational edges: staleness, contradiction, tenant isolation, retention, and deletion.

Guardrailed and recoverable execution

Explicit state machines with persisted checkpoints, retry policies matched to failure class, duplicate suppression, policy enforcement outside the model, and human escalation that resumes with full context rather than starting over.

Evaluation as an engineering practice

Outcome and side-effect assertions, efficiency metrics across steps, tokens, latency, and cost, fault injection for unavailable dependencies, and turning production traces into evaluation cases with privacy controls intact.

The Curriculum

Comprehensive Lessons! Each with theory, interactive simulation, and quiz.

The Agent Execution Loop

Governing Agent Tool Use

Planning Under Uncertainty

Memory That Earns Its Keep

Guardrailed and Recoverable Execution

Evaluating Agent Behavior

Ready to see what's really happening?

All deep dives included with your subscription. Cancel anytime.