LLM agents and automations

How to use this

• How to use this course8 min
• Scripts, workflows and agents7 min
• The virtual employee frame7 min

Exercises

• The agent that handles supportcritiquemedium4 min

Scoring

• The arithmetic of a task7 min
• The scoring frame7 min

SOP to workflow

• SOP to workflow6 min
• Definition of done and the exception list7 min

Exercises

• Which of these six tasks firstcasemedium3 min
• The cost of a mis-sent quoteestimatehard3 min
• The SOP with the decisions missingcritiqueeasy3 min

The rungs

• Rung 1 — trigger-action tools (Zapier, Make)7 min
• Rung 2 — visual orchestrators (n8n)7 min
• Rung 3 — code frameworks (LangChain, LangGraph)7 min

Choosing a rung

• Rung 4 — agent SDKs (the OpenAI Agents SDK, the Claude Agent SDK), and rung 5, the harness7 min

Exercises

• The Zapier chain with 11 stepscritiqueeasy3 min
• Pick the rung for three taskscasemedium4 min
• What the agent version costsestimatehard3 min

Graphs

• LangGraph state machines versus n8n graphs — one idea, two skins9 min
• Deterministic first: the harness pattern8 min

State and memory

• Memory and state for long-running automations7 min

Tools

• Tool design: what a tool takes, returns and refuses7 min

Exercises

• Which steps get a modelcasemedium3 min
• State for a three-day recruiting screendesignhard4 min
• The prompt that carries the whole CRMcritiqueeasy3 min

Gates and review

• Approval gates: where a human signs7 min
• Sampling review and confidence thresholds7 min

Escalation

• Escalation paths and the manager7 min

Exercises

• Gate placement for a refund flowdesignmedium3 min
• The sampling rate for support repliesestimatehard3 min
• A threshold set at 0.9 by feelcritiqueeasy3 min

Talking to systems

• How systems talk: OAuth, API keys and scopes8 min
• Webhooks versus polling7 min

Reliability

• Rate limits, retries and backoff7 min
• Idempotency and deduplication7 min

Building your own

• Backfills and building your own connector7 min

Exercises

• The retry loop that retries a 400critiqueeasy3 min
• Which failure is safe to replaycasemedium3 min
• Backfill time under a rate limitestimatemedium3 min
• A connector for a system with no webhooksdesignhard3 min

Code exercises

• Why we write these five by hand7 min

Exercises

• An idempotency key storecodeeasy3 min
• A webhook dedupercodemedium4 min
• A token-bucket rate limitercodemedium4 min
• An approval-gate state machinecodehard4 min
• A retry policycodemedium4 min

Surface and auth

• Designing the tool surface7 min
• Auth and identity for an MCP server7 min

A worked server

• A worked server: the support desk7 min

Exercises

• The tool surface for an accounting serverdesignhard4 min
• The server that exposes run_querycritiqueeasy3 min
• Read-only or read-writecasemedium4 min

Evals and monitoring

• What "working" means: the eval set7 min
• Drift and silent failure7 min

Cost

• Cost control: token budgets, caching and routing7 min
• Throughput and latency for automations7 min

Security

• Prompt injection through your own inbox7 min
• Least-privilege credentials and data leaving the tenant7 min

Exercises

• An eval set for bookkeeping prepdesignhard4 min
• The dashboard with only error countscritiqueeasy4 min
• A monthly token budget for triageestimatemedium4 min
• The inbox agent that can send mailcritiquemedium3 min
• Can this data leavecasemedium3 min

Owning a virtual employee

• Who owns a virtual employee7 min
• The weekly review8 min
• Retiring an automation7 min

Worked automations

• Inbound lead handling12 min
• Support triage11 min
• Bookkeeping prep11 min
• The content pipeline11 min
• The recruiting screen12 min
• Weekly reporting11 min

Exercises

• The review that found nothingcritiqueeasy4 min
• When to retire the lead handlercasemedium3 min
• One catalogue entry with a constraint changeddesignhard4 min