Agent, Role, Skill, Capability and Tool
The Agent Model separates who performs work, the responsibility they hold, the procedure they use, what they can accomplish and the concrete tools used to do it. Agent → Role → Skill → Capability → Tool is a conceptual sequence, not a fixed runtime object hierarchy.
This conceptual reference clarifies delegation in AI Architect Mastery's AI Driven Development Methodology. The workflow remains PRD → PLAN → TASK → IMPLEMENTATION. These distinctions neither introduce another workflow nor imply that an AAM application or technical framework is released.
Five distinct concepts
- Agent
- An AI actor performing bounded delegated development work.
- Role
- A responsibility boundary: purpose, inputs, outputs and decisions outside its scope.
- Skill
- A reusable procedure for performing a type of work.
- Capability
- What an agent can accomplish through available skills when the environment supports it.
- Tool
- A concrete mechanism used to exercise a capability.
Why this distinction matters
Assigning a role does not grant permissions. Installing a tool does not establish that an agent can use it correctly. Capability depends on the procedure, configuration, permissions, context and evidence of the intended outcome.
How to use the model when delegating
Start with the responsibility needed for a Work Item. Identify the procedure and expected output. Confirm the concrete tools and permissions in the Agent Harness, then specify evidence that demonstrates the work rather than merely tool invocation.
Constructed example: a review assignment
An agent has a reviewer role and a procedure for comparing a diff with acceptance criteria. A diff viewer is one tool. The capability to verify behavior still depends on access to suitable tests and their environment. If execution is unavailable, the agent can provide a static review but must state that runtime behavior remains unverified.
Do not assume vendor parity
A runtime may represent these concepts differently or not expose a named object for each one. Preserve the semantics without inventing native features. A role label is not a substitute for testing what the environment actually permits.