Term	Meaning
Agent	A system where the model decides what steps to take, rather than you specifying them in advance. That single property causes every hard thing about agents.
Approval fatigue	What happens when people are asked to approve too many similar things. They stop reading, and your safeguard becomes a button that adds latency.
Autonomy ladder	Five rungs from suggesting to acting freely. Choosing your rung is the most important product decision in agent design; most products should ship at rung 2 or 3.
Blast radius	The scope of damage from one bad run. Ask: what's the worst single action, how many times before anyone notices, how would we find out, how fast can we stop it.
Cap	A limit on steps, time or cost per task. Not an emergency brake, a designed feature whose triggering should produce a useful handover, not an error.
Client	In MCP, the AI application that connects to servers and makes their capabilities available to a model. Your product is the client.
Compounding error	A mistake at an early step that every later step builds on. The output looks coherent because everything after the error was internally consistent.
Give-up condition	A deliberate, designed surrender when the agent can't finish. Every agent needs one, and what it hands back is a product decision.
Irreversible action	Anything you can't undo, including anything seen by a person outside your company. Email you can delete from your system but not from their inbox.
Loop	Think, act, observe, decide, repeated until something stops it. Naming the four steps lets you diagnose which one failed.
MCP	The Model Context Protocol. An open standard, introduced by Anthropic, for how AI applications connect to external tools and data, so connections built once can be reused.
Prompt injection	Hostile instructions hidden in content the agent reads. In a chatbot the worst case is a bad answer; in an agent with write access it's an action.
Resource	In MCP, data a server exposes for reading. A resource is something; a tool does something.
Server	In MCP, the thing that exposes tools, resources and prompts. A third-party server is code with access to your systems, treat it as a supply-chain question.
Tool	A specific capability given to the agent. Your tool inventory is the complete definition of what your agent can do, which makes it product design.
Trace	The full record of a run, every tool call, argument and result. You cannot debug or evaluate an agent without traces; if you can't capture them, don't ship one.
Trajectory evaluation	Scoring the sequence of steps rather than just the final answer. Catches the agent that guessed right and the agent that took an unauthorised action but still passed.
Workflow	A fixed sequence of steps you specified. Predictable cost, enumerable paths, no flexibility. Often the right answer, and frequently mislabelled as an agent.

Two Habits That Carry Most of the Value

Choose the lowest autonomy level that delivers the value, and write the worst wrong call for every tool that writes. Both take an afternoon. Together they prevent most of the incidents and most of the projects that quietly turn out to have been workflows.

Agentic AI and MCP for Product Managers

Module 1What 'Agentic' Actually Means

Module 2How an Agent Actually Works

Module 3MCP: What It Is and Why You Should Care

Module 4Designing the Tool Layer

Module 5Permissions and the Human in the Loop

Module 6Evaluating Agents

Module 7Cost, Latency and Unpredictability

Module 8Designing What the User Sees

Module 9Risk and Blast Radius

Module 10When an Agent Is the Wrong Answer

Module 11Final Project: Design One Real Agent End to End

Module 12Appendix: The Prompt Library

Words Explained: Glossary