5/5 From Skills to Agents: What Comes Next

March 15, 2026 · 6 min read

#Claude Code #Agent SDK #skills #automation #AI #ai-augmented-workflow

Previous: Anatomy of a Plugin: Inside Marp Magic

This series started with a simple idea: you don’t need to be a developer to teach AI your workflows. You write instructions in Markdown. Claude Code follows them. The result is reusable, version-controlled, and gets better every time you refine it.

But skills have a limit: someone has to run them. You open Claude Code, type a command, review the output, provide feedback. The human is always in the loop.

What if the loop ran itself?

Phase 1: Where We Are Now

Everything in this series is Phase 1 — manual skills with human oversight:

graph LR A["Human writes skill"] --> B["Human runs skill"] B --> C["Claude processes"] C --> D["Human reviews output"] D --> E["Human refines skill"] E --> B

This is already valuable. The meeting summary skill (Post 2) saves 15 minutes per meeting. The reconciliation skill (Post 3) saves hours per month. The Marp Magic plugin (Post 4) produces entire presentations.

But you’re still the trigger. You still press the button.

Phase 2: The Agent SDK

The Claude Agent SDK (announced May 2025, available in Python and TypeScript) enables autonomous agents — programs that run skills without human intervention.

graph LR A["Trigger event"] --> B["Agent runs skill"] B --> C["Claude processes"] C --> D["Agent evaluates output"] D --> E["Agent takes action or escalates"] style A fill:#e8f5e9,stroke:#2e7d32 style E fill:#e8f5e9,stroke:#2e7d32

The key difference: the trigger is an event, not a human. A new file appears. A calendar invite is created. An email arrives. The agent runs the appropriate skill, evaluates the result, and either completes the task or escalates to a human.

The Progression

Skills you’ve built in this series sit at the foundation of a capability stack:

Level	What it is	Who triggers it	Example
Skill	Instructions in Markdown	Human, manually	Meeting summary skill
Command	Named entry point for a skill	Human, via `/command`	`/summarize-meeting`
Plugin	Package of agents + commands + skills	Human, via `/plugin:command`	Marp Magic
Agent	Autonomous program using skills	Events (file, schedule, API)	Auto-summarize calendar meetings

Each level builds on the previous. The meeting summary skill you wrote in Post 2? It becomes a component the agent calls. The reconciliation skill from Post 3? It triggers when new invoices land in a folder.

What Phase 2 Looks Like

Meeting summaries on autopilot

Trigger: Calendar event ends
→ Agent checks for attached notes/transcript
→ Runs meeting summary skill
→ Sends structured summary to attendees
→ Files it in the project folder

No human involvement. The skill does the same thing it always did — the agent just runs it automatically.

Reconciliation on new invoices

Trigger: New file in invoices/ folder
→ Agent detects new invoice
→ Runs reconciliation skill against current statement
→ If matched: logs it and moves on
→ If mismatch: sends alert to CFO with details

The CFO reviews exceptions only. Matched invoices are handled silently.

Presentation updates

Trigger: Quarterly data refresh
→ Agent pulls new statistics
→ Runs topic-researcher agent from Marp Magic
→ Updates existing presentation with fresh data
→ Runs syntax-checker and final-checker
→ Notifies presenter: "Deck updated with Q2 numbers"

The Marp Magic agents from Post 4 become components in a larger automation.

Skills Are the Foundation

This is why the skills-first approach matters. Every skill you write in Phase 1 is a building block for Phase 2:

Well-tested: You’ve refined it through iterations (Post 2’s refinement loop)
Well-documented: The skill file IS the documentation
Well-scoped: Each skill does one thing (matching the single-responsibility principle)
Version-controlled: Git tracks every change

When Phase 2 agents need to summarize a meeting, they don’t start from scratch — they load the skill you already tested and refined.

The Community

You’re not building in isolation. The Claude Code community is growing fast:

Project	Stars	What it does
claude-code-router	~25.3K	Routes tasks to specialized Claude instances
SuperClaude	~20K	Enhanced Claude Code capabilities
claude-mem	~13.1K	Persistent memory for Claude Code
awesome-claude-code	—	Curated list of plugins and resources

The skills standard published in December 2025 means skills written for Claude Code also work in Claude.ai and the Agent SDK. Write once, use everywhere.

Getting Started with Phase 2

If you’re ready to move beyond manual skills:

Master Phase 1 first — build 3-5 skills for your real workflows. Refine them. Trust the output.
Learn the Agent SDK — Building Agents with Claude Agent SDK is the official guide.
Start with scheduled triggers — a daily or weekly agent that runs your best skill on new data.
Add human-in-the-loop checkpoints — don’t go fully autonomous immediately. Have the agent draft and wait for approval.
Graduate to event-driven — file watchers, webhooks, API triggers that respond in real time.

The Series in Review

Post	What you learned
Post 1	Claude Code and skills — AI for non-developers
Post 2	Your first skill: meeting summary (hands-on)
Post 3	Real business value: account reconciliation
Post 4	Plugin architecture: 10 agents working together
Post 5	Where it’s heading: autonomous agents

The arc: learn the tools → build a skill → solve a real problem → understand the architecture → see the future.

Everything in Phase 1 prepares you for Phase 2. The skills don’t change. The trigger does.

Posts