Engineering Agentic Systems
Anyone can wire a language model to a tool and call it an agent. This free, eight-lesson advanced track is about everything that separates that demo from a production system a team can trust: engineering agentic systems, not just assembling them. You will learn to decide when a task needs an agent at all, make team standards durable in Claude Code configuration, design structured output schemas that stay honest about missing data, build tools and MCP servers other agents can use correctly, run multi-agent orchestration that survives real failures, place escalation and human review where they pay, and put agents to work headless inside CI/CD pipelines. The capstone has you design, build, and defend a small agentic system of your own. Every lesson is narrated, cited, and free. 8 lessons, about 2 hours, no signup.
What you will learn
- The architect's first question: does this task need an agent at all, and what does choosing one trade away?
- How Claude Code configuration (CLAUDE.md, scoped rules, shared skills) turns one person's judgment into a whole team's standard.
- How to design structured output schemas where "not in the document" is always a legal answer, so the model never has to invent.
- Tool and MCP server design an agent can actually use: descriptions that drive selection, errors an agent can recover from, least privilege.
- Multi-agent orchestration under real failure: briefing subagents that inherit nothing, splitting work without gaps, keeping sources attached.
- Reliability as design: escalation criteria that survive an audit, human review budgeted where it changes outcomes, independent multi-pass checks.
- Headless Claude Code in CI/CD: machine-readable output, inherited project config, and the generator-reviewer split teams trust.
Where this track sits
This is the third step of a ladder. In Building with Claude you built the parts: API calls, tool use, the agent loop. In Anatomy of an AI Agent Team you read a real multi-agent system end to end. Here you become the one making the decisions: what the model decides, what the code guarantees, and what waits for a human. The track is concept-first and code-light; if you can follow an argument about who should decide what, lesson 1 needs nothing else from you.
The lessons, in order
- Thinking like an architect Architecture is deciding where judgment lives. Learn the three trade-offs behind every agentic system before reading a single line of code.
- CLAUDE.md at team scale Configuration is how an architect's decisions become durable for a team. Read the CLAUDE.md hierarchy, scoped rules, shared commands, and plan mode.
- Schemas that refuse to lie A required field on a document that lacks the answer pressures the model to invent one. Learn to design schemas where the truth always fits.
- Tools other agents can trust An agent chooses tools by reading your descriptions. Learn to write tools, name them, design errors agents can recover from, and build an MCP server.
- Orchestration that survives contact One coordinator, many subagents, real failures. Learn to brief agents that inherit none of your context, split work, and keep sources attached.
- Reliability is a design choice Escalation criteria that work, human review where it pays, and independent multi-pass checking. Design when a person decides, not if.
- Agents in the pipeline Run Claude Code with nobody at the keyboard: headless mode, JSON output for CI, the generator-reviewer split, and review checks teams trust.
- The capstone: design, build, defend Design a small agentic system, build it in Claude Code with every pattern from this track, and write the defense that makes you an architect.
Sources and depth
Every lesson cites its sources: Anthropic's public engineering essays on building effective agents, its multi-agent research system, and tool design, plus the official Claude Code, Claude Agent SDK, and Model Context Protocol documentation. When you finish, those same sources are your next layer of depth, and the free courses in the Anthropic Academy catalog go further on every surface this track touches.