I track where the API economy is moving by reading what the providers themselves publish, and every month I pull a thread out of that firehose. Last month it was Model Context Protocol. This month I pulled the Agent Skills thread, and it tells a very different story than MCP did. Over the last 30 days — June 22 to July 23, 2026 — the providers in the API Evangelist network published far fewer posts about Agent Skills than about MCP, roughly three dozen distinct stories after I strip the cross-repo duplication (AWS reposts the same announcement across many amazon-* brands). But the volume is not the point. The maturity is.
MCP spent its first months as a gold rush and only recently hit its security reckoning. Agent Skills skipped the naive phase. The launches and the “how to build these well” conversation arrived in the same month—benchmarks, eval harnesses, and testing guides landing right next to the first provider skills. Here is the shape of that, in five themes.
If you want the durable, deduplicated version—every official provider Agent Skill, cross-linked to its provider profile and MCP server—skip the monthly snapshot and browse the directory at apis.io/skills/, where 1,489 individual skills across 75+ providers are indexed and kept current. Everything below is the month’s motion on top of that standing map.
1. The Platforms Baked Skills Into Their Clouds and IDEs
The most consequential signal is where Skills showed up: not in demos, but inside the tools developers already live in.
- AWS shipped AI Agent Skills into managed services themselves—Amazon Managed Service for Apache Flink, and Amazon MSK the month before. Skills as an operational surface for infrastructure, not just a coding aid.
- JetBrains put Agent Skills into WebStorm 2026.2, and DataGrip 2026.2 paired them with MCP tools and CLI commands for data-source management.
- Visual Studio now ships built-in Agent Skills that bring .NET and Azure expertise into the IDE directly.
- NVIDIA used Agent Skills to drive RL autoresearch workflows with NeMo, and to post-train Cosmos models in a day.
- Google introduced Agent Skills for Advertising and Measurement, and Red Hat shipped agent skills for RHEL. Salesforce is building native mobile apps in minutes with Agent Skills on Agentforce.
When Amazon, Microsoft, Google, NVIDIA, JetBrains, Red Hat, and Salesforce all ship the same primitive in a single month, it has stopped being experimental.
2. The Quality-and-Evals Conversation Arrived With the Launches
This is the theme that makes Agent Skills different from every agent primitive that came before it. Normally the “is this any good?” tooling shows up a year late. Here it arrived simultaneously.
- Arcade — SkillBench: The Quality Benchmark for Agent Skills. A benchmark, in month one.
- StackHawk — An Eval Harness for Agent Skills: How We Change Behavior on Purpose.
- Weights & Biases — How to structure effective Agent Skills and evaluate whether they actually work.
- LangChain / LangSmith — Evaluating Skills.
- Microsoft — How to test agent skills without hitting real APIs.
The lesson the industry learned from MCP—that shipping without evals is how you ship a liability—was applied to Skills on day one.
3. The Design Ethos: Say Less, Load Less
Underneath the eval tooling is a genuine point of view about what a good skill is, and it is refreshingly against maximalism.
- Zyte — The best agent skill is the one that says the least. The quote of the month.
- Google framed its launch as more autonomy, less context—the skill exists to shrink what the agent has to be told.
- The New Stack noted that the AWS Agent Toolkit ships 20+ skills, but the agent may not load the right one without a single rules file—context discipline as the real bottleneck.
- Jentic turned it around entirely: a skill that lets your agent check whether your APIs are even AI-ready.
Skills are being designed as compression, not documentation. That is the correct instinct, and it is arriving early.
4. Skills Ship Next to MCP, Not Instead of It
The providers who understand the agent surface are shipping Skills and MCP together, as complementary layers rather than competing ones.
- Hedera — Hedera MCP & Agent Skills, launched as a pair.
- Blockscout — Agent skills are live, MCP now uses the PRO API, x402 in progress.
- SigNoz shipped Agent Skills alongside its plugin surface, and Voiden brought skills for both Claude and Codex into its workflows.
This is the agent-surface layer I keep writing about: MCP for live tool access, Skills for packaged know-how, Arazzo for orchestration—shipped as a set.
5. The Long Tail of Provider Skills
Beneath the platforms, the specific, useful, single-purpose skills are proliferating.
- Hashnode — a GraphQL Agent Skill to drive the Hashnode API from Claude Code, Cursor, and Copilot.
- LambdaTest — an Appium skill for test automation with Claude Code.
- Spree — open-source eCommerce skills that turn you and your coding agent into Spree experts.
- n8n — self-improving Agent Skills built with cognee, closing the loop.
- GrowthBook — so your agents stop guessing at feature flags and experiments.
- OpenSearch ran an Agent Skills hackathon, and there is even a dedicated Spring Boot skill making the rounds.
What the Month Adds Up To
Agent Skills is a smaller surface than MCP by volume, and a more mature one by disposition. The platforms adopted it, the eval discipline arrived with the launches instead of a year later, the design ethos is compression rather than sprawl, and the smart providers are shipping it beside MCP rather than picking a side. A spec is even forming underneath, at agentskills.io, with providers like Confluent already writing skills whose only job is to review other skills against it.
The through-line for anyone building here: a skill is a contract, and contracts get tested. That is the map I keep at apis.io/skills/—1,489 real provider skills, each tied back to its provider profile and MCP server. Bookmark it, and I’ll be back next month with the motion on top.