ORA·techFrom “what is an agent” to the core design patterns (Tool Use/ReAct, Planning, Metacognition) — with interactive simulations. Adapted from Microsoft’s AI Agents for Beginners curriculum.
A new visual guide, separate from the original course below. Use it to build a mental model before entering the detailed lessons.
These are four lenses that can be combined, not a mandatory maturity ladder. Mix them according to the task’s complexity, risk, and autonomy.
Design clear instructions that guide AI toward the right kind of result in a conversation.
When to use it
How it flows
AI knows how to answer.
Supply the right information and context before AI produces an answer.
When to use it
How it flows
AI understands the right problem.
Build an orchestration layer around AI so it operates accurately, safely, and reliably.
When to use it
How it flows
AI runs safely and reliably.
Design a system in which AI plans, acts, evaluates itself, and retries until it reaches the goal or a stop condition.
When to use it
How it flows
AI completes the work autonomously.
Prompt shapes the answer · Context supplies the right data · Harness wraps it with tools, rules, and evaluation · Loop repeats with stop conditions. Combine the layers as needed.
These are the three patterns shown in the reference. Claude Code also offers Agent View and isolated worktree sessions for parallel work.
The main agent delegates a focused task to a worker with isolated context, then receives a summarized result.
Claude creates helpers immediately; each returns a separate result
More context
How to use it
Define specialists in .claude/agents/*.md; let Claude delegate automatically or invoke one explicitly with @mention.
Who controls it
The main agent delegates and synthesizes. Subagents report only to their caller and do not message each other.
When to use it
Independent work, verbose output that should stay out of the main context, or parallel research.
Multiple independent Claude Code sessions share a task list and can communicate directly with one another.
Team lead → shared task list ↔ teammates claim · update · message directly
↕ Claim work · update status
↔ Teammates message directly
More context
Agent Teams remain off by default, do not support nested teams, and teammates do not automatically get separate worktrees. Partition work by module/file to avoid overlapping edits.
How to use it
Use Claude Code 2.1.32+, enable CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1, then ask Claude to create a team and describe the roles.
Who controls it
The team lead can assign work; teammates can also claim unassigned tasks, update status, and message directly.
When to use it
Parallel investigation, competing hypotheses, or features partitioned by independent modules/files — not a sequential pipeline.
Your code uses the Agent SDK or Claude CLI to distribute, merge, evaluate, and iterate on work.
Script → multiple Workers → merged Result → repeat if needed
Fan out to N independent workers
More context
How to use it
Write an orchestrator with the Agent SDK or call claude -p --bare from scripts/CI; define limits and stop conditions yourself.
Who controls it
Your code controls loops, conditions, branches, and distribution.
When to use it
Repeatable workflows that need deterministic control, CI/background execution, branching, retries, and cost limits.
Quick choice
Start with the simplest option that works.
Sub-agent
Focused task · isolated context
Agent Team
Parallel work · direct collaboration
Custom
Repeatable workflow · code-controlled
A 5-stage curriculum with 17 lessons and 3 simulators. This section preserves the original deep-learning structure, separate from the Visual Guide above.
Unlock all 17 lessons, 3 simulators, stage quizzes, and the hands-on capstone.
Requires the Middle plan or higher
Already have an eligible plan? Sign in to unlock right away.
Not included in the current 5 stages
Coming soon
Coming soon