Kishor.Academy
Log in
← All courses

Artificial Intelligence · advanced

Agentic AI: Tools, Workflows & MCP

Design bounded tool workflows, approval gates, retry-safe actions, and MCP contracts. Test a research-and-draft assistant without sending real messages.

With Admin

advanced9 lessons · 3 sections10–15 hours of reading and lab work, plus 4–6 hours for the project; estimates exclude video.

THE TAKEAWAYS

What you’ll learn

  • Distinguish a fixed workflow from an agent whose next steps are selected by a model.
  • Model state, tool selection, stopping conditions, and a bounded execution budget.
  • Compare single-agent and multi-agent patterns without assuming more agents improve quality.
  • Explain MCP host, client, server, and protocol boundaries.
  • Define narrow tool schemas, allowlists, argument validation, and explicit error behavior.
  • Separate authentication from authorization and apply least privilege to untrusted inputs.
  • Bind approval to exact content and design retry-safe actions with a recovery plan.
  • Evaluate end-to-end traces, no-evidence paths, invalid tools, and budget exhaustion.
  • Deliver a bounded research-and-draft assistant with a production-gap checklist.

Is this for you?

Developers who understand retrieval and evaluation and want to add controlled actions to AI-assisted workflows.

Before you start

  • Complete LLM Engineering or demonstrate Python, retrieval, evaluation, and access-control fundamentals.
  • Python 3.10 or newer; familiarity with dictionaries, exceptions, and JSON.
  • Use the local simulated outbox only. No real email account, production credentials, paid API, or external publishing permission required.

PUT IT TO WORK

What you’ll build

Bounded research-and-draft assistant. Client scenario: an instructor wants a draft supported by course notes and a reviewed action. Deliverable: a bounded workflow, trace, tool contract, approval demo, and failure-case report.

Decide when a fixed workflow is enough and when model-directed choices might help. Specify tool allowlists, state, budgets, human approvals, and failure recovery before connecting external actions. Inspect a deterministic simulator that retrieves notes, drafts a response, waits for approval of exact content, and prevents duplicate outbox entries during a retry. Learn MCP host/client/server boundaries without mislabeling an ordinary Python function as a protocol server.

Course curriculum

SECTION 1Workflow design, tools, and state3 lessons
  1. Agents, workflows, and when to use themReading
  2. State, planning, and tool selectionReading
  3. Single-agent and multi-agent patternsReading
SECTION 2Permissions, approval, and protocol boundaries3 lessons
  1. MCP: hosts, clients, servers, and primitivesReading
  2. Build a narrow, well-described toolReading
  3. Authentication, authorization, and least privilegeReading
SECTION 3Failure recovery, evaluation, and a bounded assistant3 lessons
  1. Human approval, idempotency, and rollbackReading
  2. Trace, evaluate, and improve agent runsReading
  3. Capstone: a safe research-and-action agentReading