Build with judgment
Patterns are cheap to name and expensive to misuse. This course teaches when to reach for which — not a zoo of diagrams for their own sake.
// The one decision framework
Single-agent, multi-agent or feedback loop?
Ask these in order. Stop at the first “yes” that changes your architecture. Capstone work later will compose two or three — start by picking the family.
- Ask 1 · Foundations
Can one agent hold the goal and tools without thrashing?
Stay single-agent. Start with ReAct, Plan-and-Execute, Reflection or Tool Use.
- Ask 2 · Orchestration
Do you need separate skills or contexts that fight in one prompt?
Go multi-agent. Orchestrator-Subagent, Supervisor, pipelines, fan-out, MapReduce or Debate.
- Ask 3 · Feedback loops
Does quality depend on critique, retries or a human gate?
Add a feedback loop. Evaluator-Optimizer, Critic-Actor, Self-Healing or HITL.
// Journey map
Fifteen patterns. Three tiers. One decision.
Locked rows show what is coming soon — visible on the map, no dead links. Open lessons mark progress for this session only.
- Tier 1FoundationsSingle-agent loops that reason, act and check themselves0 of 4 available marked
- Tier 2OrchestrationMulti-agent shapes that divide work without losing the plot0 of 7 available marked
- Tier 3Feedback loopsFeedback loops that keep agents honest under real load0 of 0 available marked
// Tier 1 · Foundations
Foundations
Reach for this tier when: One agent can hold the goal and tools without thrashing.
- Reflection / Self-Critique
When a first draft is cheap and a second pass catches real mistakes.
~11 min
- Tool Use / Function Calling
When the model must call structured tools instead of inventing facts.
~13 min
// Tier 2 · Orchestration
Orchestration
Reach for this tier when: No single agent can hold all skills or context cleanly.
- Hierarchical Agents
When goals nest: strategy → tactics → tools, with clear ownership per layer.
~13 min
- Sequential Pipeline
When stages are fixed and each step’s output is the next step’s only input.
~11 min
- Parallel Fan-Out / Fan-In
When independent sub-tasks can run at once and merge into one answer.
~12 min
// Tier 3 · Feedback loops
Feedback loops
Reach for this tier when: Quality depends on critique, retries or a human gate.
- Coming soon
Evaluator-Optimizer
When a scorer can reject work until an optimizer improves it.
- Coming soon
Critic-Actor
When a dedicated critic reviews every actor draft before it ships.
- Coming soon
Self-Healing / Retry Loop
When failures are expected and a bounded retry with diagnosis beats giving up.
- Coming soon
HITL (Human-in-the-loop)
When irreversible or high-stakes actions need a human gate.
Open a lesson, then mark “Got it” to fill the map.
Progress stays while you move around these pages. Nothing is saved.
// Learn by doing
Put it to work
- ReAct →Foundations: reason, act, observe, stop.
- Plan-and-Execute →Foundations: plan once, execute in order, replan on failure.
- Reflection →Foundations: draft, critique against a rubric, revise.
- Tool Use →Foundations: schema-bound tools, no invented payloads.
- Orchestrator-Subagent →Orchestration: dispatch specialists, merge results.
- Supervisor →Orchestration: route, score, and veto — do not draft.
- Hierarchical Agents →Orchestration: strategy → tactics → tools, no rank-skipping.
- Sequential Pipeline →Orchestration: fixed stages, typed handoffs only.
- Parallel Fan-Out →Orchestration: independent branches, then join.
- MapReduce →Orchestration: identical map jobs, then reduce.
- Debate →Orchestration: bounded pro/con, then a judge.
- Tier 1 quiz →Four questions on the decision framework and ReAct.
- Tier 2 quiz →Four questions on Supervisor, fan-out, MapReduce, and Debate.
- Cheat sheet →All fifteen patterns, one “reach for when” line each.
Coming soon: four more lessons, Tier 3 quiz, and the capstone.
How this course stays honest
No fake benchmarks. Mini-projects are clearly stubbed. Progress is session-only — nothing is saved. Locked syllabus rows are visible so you can see the map; they do not deep-link to missing pages.
This course teaches pattern practice. Mir's studio fleet does not claim production eval suites or regression harnesses for every agent — the lessons are the map, not a claim that every live agent ships with matching eval coverage.