Architectural Layers

Five orthogonal concerns that compose into a complete agent

L1 Tools & Execution

What the agent CAN do. The foundation: tools give the model capabilities to interact with the world.

s01tools

The Agent Loop

Bash is All You Need

81 LOC1 tools

The entire agent is a while loop + one tool

s02tools

Tools

The Loop Didn't Change

115 LOC4 tools

Adding tools means adding handlers, the loop stays the same

L2 Planning & Coordination

How work is organized. From simple todo lists to dependency-aware task boards shared across agents.

s03planning

TodoWrite

Plan Before You Act

168 LOC5 tools

Visible plans improve task completion and accountability

s04planning

Subagents

Fresh Context via Task Tool

141 LOC5 tools

Process isolation = context isolation

s05planning

Skills

SKILL.md + tool_result Injection

172 LOC5 tools

Skills inject via tool_result, not system prompt

s07planning

Tasks

Persistent CRUD with Dependencies

197 LOC8 tools

File-based state survives context compression

L3 Memory Management

Keeping context within limits. Compression strategies that let agents work infinitely without losing coherence.

s06memory

Compact

Strategic Forgetting

195 LOC5 tools

Forgetting old context enables infinite-length sessions

L4 Concurrency

Non-blocking execution. Background threads and notification buses for parallel work.

s08concurrency

Background Tasks

Fire and Forget

188 LOC6 tools

Non-blocking daemon threads + notification queue

L5 Collaboration

Multi-agent coordination. Teams, messaging, and autonomous teammates that think for themselves.

s09collaboration

Agent Teams

Teammates + Mailboxes

343 LOC10 tools

Persistent teammates with async mailbox inboxes

s10collaboration

Team Protocols

Shutdown + Plan Approval

405 LOC12 tools

Same request-response pattern, two applications

s11collaboration

Autonomous Agents

Idle Cycle + Auto-Claim

483 LOC14 tools

Polling + timeout makes teammates self-organizing