One curriculum, three stacks
Agent Workflows, Skills & Multi-Agent Systems
Learn how production agents are actually put together — workflows vs autonomous loops, reusable skills, multi-agent topologies, and context engineering — with every example implemented for Gemini, Claude and OpenAI side by side.
Concepts first
Each lesson teaches the pattern before any SDK: workflows, skills, multi-agent topologies, context budgets, tool contracts.
Three stacks, one example
Every code sample is written for Gemini, Claude and OpenAI. Switch the tab once — the whole track remembers your stack.
Interview-mapped
Every lesson ends with the questions it prepares you for, and links to the case studies where you can practise them.
The curriculum
Read in order the first time; each lesson stands alone afterwards.
- 1Foundations14 min read
Agent Workflows vs Autonomous Agents: The Five Patterns You Actually Ship
Most "agents" in production are workflows: model calls arranged in code-defined paths. Learn the five patterns — chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer — and when to hand control to a real loop.
- ✓Explain the difference between a workflow and an agent in one sentence an interviewer will accept
- ✓Pick the right pattern for a task and justify it with cost, latency and reliability arguments
- ✓Implement a routing workflow with the Gemini, Claude and OpenAI SDKs
Open lesson →
- 2Foundations12 min read
Agent Skills: Packaging Expertise So Any Model Can Load It On Demand
A skill is a folder of instructions, scripts and references an agent loads only when a task needs it. Learn the SKILL.md format, progressive disclosure, and how to make one skill work across Claude, Gemini and OpenAI.
- ✓Write a SKILL.md with a description that triggers reliably and a body that stays out of context until needed
- ✓Explain progressive disclosure and why it beats one giant system prompt
- ✓Load the same skill into Claude, Gemini and OpenAI agents
Open lesson →
- 3Intermediate16 min read
Multi-Agent Systems: Orchestrators, Handoffs and When One Agent Is Enough
Sub-agents, supervisors, handoffs and swarms — what each topology buys you, what it costs, and the same research team implemented with Google ADK, the Claude Agent SDK and the OpenAI Agents SDK.
- ✓Choose between supervisor, handoff and parallel topologies and defend the choice
- ✓Explain why sub-agents are primarily a context-isolation tool, not a "smarter" architecture
- ✓Build an orchestrator with two workers in ADK, Claude Agent SDK and OpenAI Agents SDK
Open lesson →
- 4Intermediate15 min read
Context Engineering: Budgeting, Caching and Compacting the Window
The context window is the agent’s working memory and its bill. Learn what belongs in it, how to budget it, how prompt caching works on each provider, and how to compact a long-running agent without losing the plot.
- ✓Draw the layers of an agent’s context and assign each a token budget
- ✓Use prompt/context caching on Gemini, Claude and OpenAI and explain what each caches
- ✓Implement compaction for a long-running agent and know what must survive it
Open lesson →
- 5Intermediate13 min read
Tools and MCP Across Providers: One Server, Three Clients
Function calling looks different in each SDK but is the same protocol underneath. Learn the request/response shapes side by side, then connect a single MCP server to Gemini, Claude and OpenAI.
- ✓Translate a tool definition and a tool-call round-trip between the three SDKs from memory
- ✓Connect one MCP server to all three providers
- ✓Explain what MCP standardises and what it deliberately leaves to you
Open lesson →
- 6Advanced11 min read
Gemini vs Claude vs OpenAI: Choosing an Agent Stack and Talking About It in Interviews
A vocabulary map across the three ecosystems — SDKs, agent frameworks, coding harnesses, skill and context conventions — plus a decision framework and the questions interviewers ask about provider choice.
- ✓Map any concept (skills, sub-agents, caching, MCP) to its name in each ecosystem
- ✓Argue for a provider choice on engineering grounds rather than benchmark headlines
- ✓Design for portability so a provider switch is a week, not a rewrite
Open lesson →
- 7Intermediate18 min read
Hands-On: Build a Claude Code Plugin, an OpenAI App, and a Gemini Agent
Three walkthroughs, one weekend: package skills, commands and MCP servers into a Claude Code plugin; ship a ChatGPT app (and Codex plugin) on top of an MCP server; build and deploy a Gemini agent with ADK. Each ends with a shippable artifact.
- ✓Scaffold, test and distribute a Claude Code plugin from a git repo
- ✓Turn an MCP server into a ChatGPT app with a rendered UI, and reuse it as a Codex plugin
- ✓Build a Gemini ADK agent with tools and sessions, run it locally, and deploy it to Vertex AI Agent Engine
Open lesson →
- 8Advanced15 min read
Comparing Models: Parameters, Benchmarks, and What Actually Predicts Agent Performance
Parameter counts are undisclosed for every frontier model and misleading for the open ones. Learn what each headline benchmark measures, how leaderboards get gamed, how the Gemini, Claude and OpenAI line-ups are tiered, and how to run the only eval that matters — yours.
- ✓Explain why "how many parameters" is the wrong first question, and what to ask instead
- ✓Read a benchmark table critically — know what SWE-bench, GPQA, τ-bench, AIME, HLE and Arena scores each measure and how they fail
- ✓Choose a model tier per workload with a cost/latency/quality argument, and build a 50-case eval to confirm it
Open lesson →
Practise it under interview conditions
The AI Agent Orchestration case study asks you to design exactly these systems — durable runs, budget governors, approval gates — with an interviewer scorecard.