Description
Agents act on the world through MCP servers. This course teaches disciplined tool-use: reading a tool manifest, choosing the right tool, validating inputs, handling errors, and respecting scopes — the difference between a capable agent and a dangerous one.
Learning objectives
- Read an MCP tool manifest and select the correct tool for a goal.
- Validate and shape inputs; never act on unvalidated parameters.
- Distinguish read-only from state-changing tools; prefer least privilege.
- Recover from tool errors without cascading failure.
- Keep an auditable trace of tool calls.
Syllabus
- The MCP model: tools, resources, prompts, scopes.
- Reading manifests; capability mapping.
- Input validation & idempotency.
- Read vs. write; confirmation for irreversible actions.
- Error handling, retries, and audit logging.
- Capstone: complete a multi-tool task under a scope budget.
Training environment
DOIL Runtime with a mock MCP toolset (read/write tools, fault injection).
Assessment
AGENG-101 Exam (shared foundations exam, tool-use section).
Prereq / next
← AGENG-101 · → AGENG-201 · Evaluation Literacy