a

a.md

The Agent SDK Click: A Brief Report

The Metaphor

Claude Code is a brilliant generalist contractor. It can do plumbing, electrical, framing, painting. But generalists have a failure mode: they get distracted, forget what they were doing, sometimes rip out work they did yesterday.

Agents are specialists you can call in. An electrician shows up, does the wiring, leaves. They don't touch your plumbing. They don't have opinions about your paint colors. They do ONE thing and return a result.

The Killer App for You

You already discovered it: your manager/sub-agent workflow. The Agent SDK just formalizes and packages those sub-agents.

Instead of:

• Setting up a new Claude Code session manually
• Writing GitHub issues to coordinate
• Checking status files

You write a Python/TypeScript agent ONCE: @agent("blender-rig-tweaker") def tweak_rig(blend_file: str, instruction: str) -> str:

Agent has Blender CLI access

Iterates: modify script → render frame → check result

Returns: path to modified .blend file

Then Claude Code (your interactive session) just calls it: "Use the blender-rig-tweaker agent to make this walk cycle bouncier"

Why This Matters for Creative Work

Creative tools have tight feedback loops if you know how to script them: ┌─────────┬─────────────────────────┬──────────────────────────────────────────┬──────────────────────┐ │ Tool │ Input │ Agent Can... │ Verification │ ├─────────┼─────────────────────────┼──────────────────────────────────────────┼──────────────────────┤ │ Blender │ .blend + instruction │ Modify Python rig script, render preview │ Compare frames │ ├─────────┼─────────────────────────┼──────────────────────────────────────────┼──────────────────────┤ │ Roblox │ .lua + desired behavior │ Edit script, run in Studio │ Check output logs │ ├─────────┼─────────────────────────┼──────────────────────────────────────────┼──────────────────────┤ │ Bitwig │ .bwproject + vibe │ Modify XML, render audio │ Analyze spectrum │ ├─────────┼─────────────────────────┼──────────────────────────────────────────┼──────────────────────┤ │ Music │ MIDI + mood │ Generate variations │ Compare to reference │ └─────────┴─────────────────────────┴──────────────────────────────────────────┴──────────────────────┘ The agent handles the iterate until right loop autonomously while your main Claude Code session stays clean and high-level with you.

When to Reach for Agent SDK

Reach for it when:

1. You've done a task 3+ times manually in Claude Code
2. The task has a verifiable end state (renders, compiles, passes tests, sounds right)
3. You want Claude Code to delegate not do

Don't reach for it when:

• The task requires your taste in real-time
• You need to explore/discover, not execute
• It's truly one-off

The One-Liner

Agents are encapsulated expertise that Claude Code calls like a function—so it can stay creative with you instead of getting lost in the weeds.