Agents, Skills, Tools, and MCP
Separate agent decisions, reusable methods, deterministic actions, workflow state, and external connections.
“Agent Basics for Product Managers” part five
Previous: Context, Sessions, and Memory · Series home · Next: Agent Autonomy
Agent products often use the words agent, skill, tool, MCP, workflow, and prompt in the same conversation. They all contribute to task completion, but they belong to different layers.
The shortest useful distinction
| Object | Responsibility | PRD example |
|---|---|---|
| Agent | Decide what to do next and own progress | Decide whether to clarify the BRD or draft the PRD |
| Skill | Define how a recurring task is performed | A PRD review method and quality rubric |
| Tool | Perform one deterministic action | Read a document, create a patch, save a review |
| Workflow | Control critical steps and state transitions | Extract→clarify→draft→review→approve |
| MCP | Connect a host to external capabilities through a protocol | Connect a document store or task system |
| Instructions | Define role, principles, and output constraints | Never present an assumption as a fact |
The agent selects a skill. The skill uses tools according to a method. The workflow controls the critical path. MCP connects capabilities.
A skill is not an impressive prompt
A useful skill is an executable work instruction: when it should trigger, which inputs it needs, what context to load, which steps and tools to use, what to produce, how to judge quality, and what to do when information is missing.
“Expert PRD reviewer” is a role. “Review evidence, actors, primary flow, edge cases, and acceptance criteria, then report blockers against a rubric” begins to define a skill.
Tools are the boundary of real action
A model saying “I updated the document” changes nothing. Reality changes only when the runtime invokes a write tool and receives a successful result.
A production tool needs a narrow purpose, input and output schemas, server-side identity and permission checks, declared side effects, timeout and idempotency behavior, and an audit event. Narrow tools such as propose_document_patch and apply_approved_patch are easier to authorize and test than manage_project.
MCP connects; it does not authorize
MCP makes tools, resources, and prompts interoperable across hosts and servers. A successful connection does not make a server trustworthy or grant the current user every action. Business identity, scope, parameter limits, and audit remain product responsibilities.
Product exercise
Split one feature into four columns: agent decisions, skill method, workflow state, and tool actions. Put each capability in one column first. If it appears in three or four, the responsibility is still unclear.
Layering is what lets a team improve methods, actions, permissions, and quality independently.