Learn Claude Code

Learning Path

s01 to s17: Progressive Agent Harness Design

Layer Legend

Tools & Execution
Planning & Control
Memory Management
Concurrency & Scheduling
Multi-Agent Platform
01
s01Minimal model/tool loop

The Agent LoopOne Loop Is All You Need

106 lines of code1 tools

The smallest useful agent is a loop that calls the model, runs tools, and feeds results back.

Learn More
02
s02Tool dispatch map

Tool UseAdd a Tool, Add Just One Line

145 lines of code5 tools

The loop stays stable while capabilities register into a dispatch table.

Learn More
03
s03Permission gate

PermissionCheck Permissions Before Execution

181 lines of code5 tools

Dangerous actions need a harness decision point before the shell runs.

Learn More
04
s04Lifecycle hooks

HooksHang on the Loop, Don't Write into It

203 lines of code5 tools

Cross-cutting behavior belongs around the loop, not tangled inside it.

Learn More
05
s05Todo manager

TodoWriteAn Agent Without a Plan Drifts Off Course

280 lines of code6 tools

Explicit plans keep long-running work visible and correctable.

Learn More
06
s06Isolated subtask context

SubagentBreak Large Tasks into Small Ones with Clean Context

287 lines of code6 tools

Subagents give each subtask a clean message history while preserving the main thread.

Learn More
07
s07On-demand skill loader

Skill LoadingLoad Only When Needed

302 lines of code6 tools

Inject specialized knowledge only when the task actually needs it.

Learn More
08
s08Context compaction

Context CompactContext Will Fill Up

423 lines of code5 tools

Compression keeps the conversation usable when the context window gets crowded.

Learn More
09
s09Durable memory layer

MemoryKeep a Layer That Doesn't Lose Details

669 lines of code5 tools

Some facts should survive summarization and future sessions.

Learn More
10
s10Task board

Task SystemBreak Big Goals into Small Tasks

421 lines of code10 tools

A task graph turns vague goals into ordered, observable work.

Learn More
11
s11Background execution

Background TasksSlow Operations Go to the Background

401 lines of code5 tools

The agent can keep reasoning while slow work completes elsewhere.

Learn More
12
s12Scheduled task creation

Cron SchedulerProducing Work on a Schedule

639 lines of code5 tools

Recurring work should be created by the harness, not remembered by the model.

Learn More
13
s13Team runtime with task-bound worktrees

Agent Team RuntimePersistent Teammates, Atomic Claims, Task-Bound Worktrees

1525 lines of code5 tools

Persistent teammates can reliably discover and execute parallel work when the runtime owns messaging, atomic claims, and task-bound working directories.

Learn More
14
s14MCP tool bridge

MCP ToolsExternal Tools, Standard Protocol

441 lines of code5 tools

External services can become agent tools through a standard discovery and call protocol.

Learn More
15
s15Integrated harness

Integrated HarnessMany Mechanisms, One Loop

2618 lines of code25 tools

The integrated harness is still one loop, surrounded by the systems introduced across the course.

Learn More
16
s16Resumable workflow runtime

Workflow RuntimeScripts Own Fixed Orchestration

722 lines of code26 tools

When orchestration has a fixed shape, code can make it parallel, deterministic, and resumable.

Learn More
17
s17Goal completion gate

Goal LoopIndependent Evaluation Decides When to Stop

790 lines of code5 tools

A durable goal keeps the loop working until an independent evaluator finds the completion condition satisfied in the conversation.

Learn More

LOC Growth

s01
106
s02
145
s03
181
s04
203
s05
280
s06
287
s07
302
s08
423
s09
669
s10
421
s11
401
s12
639
s13
1525
s14
441
s15
2618
s16
722
s17
790