Learn Claude Code

An agent is just a loop: send messages, execute tools, feed results back, repeat.

Adding a tool means adding one handler. The loop never changes.

A visible plan keeps the agent on track when tasks get complex.

A subagent is mainly a context boundary, not a process trick.

Discover cheaply, load deeply -- only when needed.

Compaction isn't deleting history -- it's relocating detail so the agent can keep working.

Safety is a pipeline, not a boolean: deny, check mode, allow, then ask.

The loop owns control flow; hooks only observe, block, or annotate at named moments.

Memory gives direction; current observation gives truth.

The model sees a constructed input pipeline, not one giant static string.

Most failures aren't true task failure -- they're signals to try a different path.

Todo lists help a session; durable task graphs coordinate work that outlives it.

Background execution is a runtime lane, not a second main loop.

Scheduling is not a separate system -- it just feeds the same agent loop from a timer.

Teammates persist beyond one prompt, have identity, and coordinate through durable channels.

A protocol request is a structured message with an ID; the response must reference the same ID.

Autonomy is a bounded mechanism -- idle, scan, claim, resume -- not magic.

Tasks answer what; worktrees answer where. Keep them separate.

External capabilities join the same routing, permission, and result-append path as native tools.