Course / Level 3
Level 3
Actionable
An agent can execute defined actions — booking, quoting, querying — against your systems.
Level 3 is the line where engineering shows up. It also splits in two: being callable (3a) through a documented API contract, and being tool-exposed (3b) through standardized agent tool interfaces like MCP. The actions you expose are scoped, observed, authorized server-side, and reversible — and for anything destructive, a human stays in the loop.
This is the first level that makes someone on your side responsible for an integration. We cover both the contract that makes calls reliable and the tool surface that lets agents discover and invoke them safely.
This level’s lessons are part of the course.
These lessons need an active enrollment. One payment unlocks every level — and Level 0 stays free for everyone. Already enrolled? Log in to pick up where you left off.
7 steps in this level
-
Step 1
A machine-readable API contract
OpenAPI 3.1+ or a GraphQL schema agents can read to know exactly what they can call.
-
Step 2
Idempotency & structured errors
Idempotency keys for mutations, RFC 9457 errors with retryability, and async terminal states.
-
Step 3
Auth & rate limiting
OAuth 2.1 on the API and rate limiting that speaks Retry-After instead of just failing.
-
Step 4
An MCP server with annotated tools
A Model Context Protocol server whose tools carry the behavioral annotations agents rely on.
-
Step 5
Agent cards & skills
A2A agent-card discovery and Agent Skills bundled as self-contained capabilities.
-
Step 6
Tool safety
Server-side authorization on every call, human-in-the-loop for destructive actions, and tool-poisoning defenses.
-
Step 7
Recap
What changed, why it matters, and how to package this level into a service you can sell.