ModernCS
Join for free
LangChain

Generative AI · Taught with LangChain

AI Agents and Tool Use

Tool calling, MCP servers, the agent loop, and the permissions around it. You build agents that finish the task instead of looping forever, and you can show exactly what they did.

Download the course plan
  • Advanced
  • 4 phases
  • 22 sessions
  • First session free

Syllabus

4 phases · 22 sessions · each phase ends in something you have built

1

The Loop, By Hand

Understand what an agent is before a framework hides it from you.

$108

this phase

  1. 1.1

    The Loop Is Ten Lines

    Free preview

    Write the while loop yourself: model call, tool call, append the result, check the stop condition.

    Open
  2. 1.2

    Tool Schemas That Work

    Naming, descriptions, enums, and the JSON Schema fields a model actually pays attention to.

  3. 1.3

    Errors the Model Can Read

    Bad arguments, hallucinated tool names, and error text written for a model instead of a human.

  4. 1.4

    Plan or React

    ReAct, upfront plans, replanning, and the tasks where planning costs more than it saves.

  5. 1.5

    Budgets and Stop Conditions

    Turn caps, wall clock, token and dollar ceilings enforced in code, not asked for in the prompt.

By the end of this phase:A framework-free agent loop with a turn cap, a dollar cap, and a trace you can read line by line.
2

Tools and the MCP Layer

Connect an agent to real systems over MCP without trusting whoever wrote the tool.

$108

this phase

  1. 2.1

    MCP, Concretely

    Tools, resources, streamable HTTP, and what the 2026-07-28 revision moved into extensions.

  2. 2.2

    Writing a Server

    Build a server with typed input and output schemas, then drive it from two different clients.

  3. 2.3

    Auth Without Confused Deputies

    OAuth 2.1, protected resource metadata, and the audience check that stops token passthrough.

  4. 2.4

    Tool Poisoning and Rug Pulls

    Instructions hidden in tool descriptions, line jumping at tools/list, and pinning what you approved.

  5. 2.5

    Schema Drift

    Hash the whole tool surface, diff it on every deploy, and catch the silent breaking change.

  6. 2.6

    Too Many Tools

    150,000 tokens of definitions down to 2,000: progressive disclosure and code execution over MCP.

By the end of this phase:An MCP server of your own with OAuth, plus a client whose tool surface is hashed and pinned.
3

State That Survives

Give the agent state that outlives the process and memory that earns its tokens.

$108

this phase

  1. 3.1

    Why a Graph at All

    create_agent, StateGraph, and the point where a hand-rolled loop stops being worth keeping.

  2. 3.2

    Checkpointers and Threads

    State reducers, a Postgres checkpointer, and a thread id you own rather than one you invent.

  3. 3.3

    Long-Term Memory

    Namespaced writes to a store, scoped retrieval, and a TTL so stale facts stop coming back.

  4. 3.4

    Context Rot

    Degradation measured at 100k tokens, schema-driven compaction, and what the second summary dropped.

  5. 3.5

    Durable Resume

    Checkpoints are not durable execution: idempotency keys, at-least-once steps, replay after kill -9.

By the end of this phase:A long-running agent killed mid-run and resumed without repeating a single side effect.
4

Blast Radius and Bills

Run the thing somewhere it can be watched, stopped, and paid for.

$108

this phase

  1. 4.1

    Sandboxing for Real

    Shared-kernel containers against Firecracker microVMs, and egress rules that matter more than both.

  2. 4.2

    Permissions and Capabilities

    Read-only by default, scoped credentials, and the Replit agent that dropped a production database.

  3. 4.3

    Approval a Human Can Give

    interrupt(), a diff a reviewer can actually judge, and what happens when nobody clicks for a week.

  4. 4.4

    Tracing the Whole Run

    OpenTelemetry GenAI spans, run and turn hierarchy, and finding the step that actually broke.

  5. 4.5

    The Runaway Drill

    Build a two-agent clarification loop on purpose, watch the spend curve, then stop it server side.

  6. 4.6

    When Not to Multi-Agent

    Supervisor and handoff patterns, a 15x token bill, and the equal-budget single-agent baseline.

By the end of this phase:A sandboxed agent with approval gates, full traces, and a spend cap proven by a runaway drill.

Tools you will use

  • LangGraph 1.2
  • LangChain 1.3
  • Model Context Protocol 2026-07-28
  • MCP Python SDK
  • Pydantic AI 2.x
  • LangSmith
  • Langfuse
  • OpenTelemetry GenAI conventions
  • E2B / Modal sandboxes
  • Temporal

What you will build

  • The Bounded Loop

    Hand-written agent loop, hard caps, a readable trace

  • Your Own MCP Server

    OAuth, pinned schemas, and a rug-pull test that fails loudly

  • The Agent That Survives

    Killed mid-run, resumed, sandboxed, traced, under budget