Hey
I built two Agent Skills last week.
The first one turns rough text into polished frontend code using React and Tailwind CSS — drop in some ideas or a sketch, get back a website in seconds. The second one reads all my writing samples and generates a style guide so that anything my assistant writes next sounds just like me.
Both took about 20 minutes. Just markdown files with clear instructions.
Here's what a snippet of the 150 line voice-analyzer skill looks like:
# Voice Analyzer Skill
When given writing samples:
1. Extract sentence patterns (length, rhythm, punctuation)
2. Identify vocabulary preferences (technical vs. casual)
3. Note structural habits (how paragraphs open/close)
4. Output a style guide with 3-5 concrete rules
Use this guide when writing future content.They're simple. Markdown Files. No Infrastructure. No Loops. No Orchestration. Chef's kiss.
And the whole time I kept thinking: this is exactly what we've been building in Agent Studio.
The industry finally gets it: skills over agents
Here's what I want to unpack.
For the past few years, the default mental model was "build AI agents." Need to handle expenses? Build an expense agent. Need to onboard employees? Build an onboarding agent. Every domain gets its own agent.
But that's not actually how it works best.
What Claude Skills does — and what Agent Studio has always done — is different: you build tools, not agents. You give one powerful assistant domain-specific skills, and it figures out when to use them.
My frontend design skill doesn't "think" about design. It's a set of instructions that my Agent follows when I ask for UI help. My voice-analyzer skill doesn't "understand" writing — it's a structured process that extracts patterns from samples.
Same deal with Agent Studio plugins. You're not building separate AI brains. You're teaching one assistant how to handle specific tasks — from simple lookups to multi-turn conversations with approvals and validation.
The mental model shift
This matters because the alternative — orchestrating multiple agents talking to each other — is something the industry is still figuring out. Recent research from Berkeley (MAST) found that even state-of-the-art multi-agent frameworks fail 60-70% of the time due to coordination issues. It's not that multi-agent is inherently bad — it's that no one has cracked it yet.
The honest take on Agent Skills
They're simple. Markdown files. No infrastructure. Claude Code can even generate new skills itself.
But here's where it's different: Agent Skills are built for developers who live in their terminals. It's a different platform, catered toward developers working in their IDE.
Agent Studio plugins are built for enterprise teams who hit walls that markdown instructions can't solve:
- Ambiguous user input at scale — when "create a ticket for John" could match 47 Johns, dynamic resolvers filter to valid options before the user ever sees a list
- Policies that can't be trusted to interpretation — typed slots enforce "due date must be in the future" at the system level, not the prompt level
- Workflows that need human checkpoints — approval steps baked into the process, not bolted on after
- Automation that runs without a user present — triggers from webhooks, schedules, and events that kick off plugins autonomously
- Compliance requirements — audit trails that log every decision, every approval, every action taken
Same philosophy. Different depth.
Tradeoff: Claude Skills are faster to build but lack audit trails and approval workflows. Agent Studio plugins take longer to set up but give you enterprise controls out of the box.
Why does this convergence matter?
Anthropic building Skills is exciting. It's the same philosophy we've been investing in — and it means this pattern is becoming industry standard. More tools adopting this approach makes the whole ecosystem stronger.
The industry is aligning: don't build swarms of agents. Build skills. Build plugins. Teach your assistant.
If you're already building in Agent Studio, you're ahead. The patterns you're learning — slots, processes, actions, triggers — are the same patterns showing up everywhere now.
And if you're experimenting with Claude Skills on the side (like I am), bring those ideas back. The mental model transfers.
Quick hits:
→ Developer Office Hours are dropping back this month. After a great run last year, we're kicking off 2026 with more live builds, Q&A, and plugin teardown. Join the community for announcements.
→ LDAP actions are being shipped to Agent Studio. If you've been wiring up directory lookups the hard way, this one's for you. Check the docs out!
Worth reading:
- Equipping agents for the real world with Agent Skills — Anthropic's engineering deep dive on Skills architecture. Key insight: "Skills are not agents—they're structured instructions that agents invoke." Progressive disclosure, bundled scripts, the whole philosophy.
- Human data will be a $1 trillion/year market — Ali Ansari makes the case that as automation scales, human judgment becomes more valuable, not less. Every function you automate still needs human expertise to define "good." That's not annotation — it's structured judgment.
That's it. See you in two weeks.
– Kevin
DevRel @ Moveworks | Agent Studio
P.S. We're dropping some video shorts on plugin patterns soon. Reply if you want early access, or join the community to catch them first: community link
Subscribe to the newsletter to get it direct to your inbox here: aka.mw/newsletter