How to Build an AI Agent
How to Build an AI Agent

Lernpfad

s01 bis s12: Fortschreitendes Agent-Design

Schichtenlegende

Tools & Execution
Planning & Coordination
Memory Management
Concurrency
Collaboration
01
s01Single-tool agent loop

The Agent LoopBash is All You Need

84 Zeilen Code1 Tools

The minimal agent kernel is a while loop + one tool

Mehr erfahren
02
s02Tool dispatch map

ToolsOne Handler Per Tool

120 Zeilen Code4 Tools

The loop stays the same; new tools register into the dispatch map

Mehr erfahren
03
s03TodoManager + nag reminder

TodoWritePlan Before You Act

176 Zeilen Code5 Tools

An agent without a plan drifts; list the steps first, then execute

Mehr erfahren
04
s04Subagent spawn with isolated messages[]

SubagentsClean Context Per Subtask

151 Zeilen Code5 Tools

Subagents use independent messages[], keeping the main conversation clean

Mehr erfahren
05
s05SkillLoader + two-layer injection

SkillsLoad on Demand

187 Zeilen Code5 Tools

Inject knowledge via tool_result when needed, not upfront in the system prompt

Mehr erfahren
06
s06micro-compact + auto-compact + archival

CompactThree-Layer Compression

205 Zeilen Code5 Tools

Context will fill up; three-layer compression strategy enables infinite sessions

Mehr erfahren
07
s07TaskManager with file-based state + dependency graph

TasksTask Graph + Dependencies

207 Zeilen Code8 Tools

A file-based task graph with ordering, parallelism, and dependencies -- the coordination backbone for multi-agent work

Mehr erfahren
08
s08BackgroundManager + notification queue

Background TasksBackground Threads + Notifications

198 Zeilen Code6 Tools

Run slow operations in the background; the agent keeps thinking ahead

Mehr erfahren
09
s09TeammateManager + file-based mailbox

Agent TeamsTeammates + Mailboxes

348 Zeilen Code10 Tools

When one agent can't finish, delegate to persistent teammates via async mailboxes

Mehr erfahren
10
s10request_id correlation for two protocols

Team ProtocolsShared Communication Rules

419 Zeilen Code12 Tools

One request-response pattern drives all team negotiation

Mehr erfahren
11
s11Task board polling + timeout-based self-governance

Autonomous AgentsScan Board, Claim Tasks

499 Zeilen Code14 Tools

Teammates scan the board and claim tasks themselves; no need for the lead to assign each one

Mehr erfahren
12
s12Composable worktree lifecycle + event stream over a shared task board

Worktree + Task IsolationIsolate by Directory

694 Zeilen Code16 Tools

Each works in its own directory; tasks manage goals, worktrees manage directories, bound by ID

Mehr erfahren

LOC-Wachstum

s01
84
s02
120
s03
176
s04
151
s05
187
s06
205
s07
207
s08
198
s09
348
s10
419
s11
499
s12
694