Twitter/X AI + Agentic Age: Skills Become Infrastructure.

Dates: March 15-28, 2026

Source note: I filtered my X bookmarks by tweet postedAt date. The bookmarks API gives the tweet date, not the exact date I bookmarked it, which is an annoying but useful distinction.

This was the fortnight when “skills” stopped sounding like a Claude feature and started looking like a distribution format. OpenAI shipped Codex subagents. Vercel bundled 47-plus specialized skills into a plugin. Figma opened use_figma MCP and skills. Ramp released a CLI for company finance. Factory launched Missions.

The shape was suddenly clear: the new app store is not an app store, because of course that would be too normal. It is a pile of CLIs, MCP servers, slash commands, markdown instruction packs, plugins, registries, and half-standardized directories called things like .agents/skills.

The interesting part is not that agents can use tools. The interesting part is that tools are being rewritten so agents can use them without wasting half the context window eating the menu. That is why Ramp’s “fewer tokens than MCP” claim mattered, why Cloudflare’s Markdown docs mattered, and why Vercel’s plugin push mattered. The industry was discovering that an agent does not need a brochure. It needs an affordance.

Key movements

Subagents became normal. Codex shipped subagents, and Claude, Factory, and Ramp patterns all pointed toward parallel specialized work rather than a single chat thread wandering through a repo with a lantern. This is not just a productivity claim. Once work is split across subagents, the product has to answer how goals are divided, how results are merged, and how a human reviews the combined output without needing a second job.

Skills became infrastructure. Vercel, Figma, shadcn, Ramp, and individual builders turned knowledge into installable agent capability. The form matters because a skill is neither documentation nor code in the usual sense; it is an instruction package that lets an agent inherit a local way of working.

Design moved into the agent loop. Figma MCP, Stitch, shadcn presets, Emil Kowalski’s design engineering skill, and Cheng Lou’s Pretext thread all circled the same uncomfortable fact: agents can now produce a lot of software very quickly, and much of it looks like it was assembled by a committee of defaults. Taste became a bottleneck, which is annoying because taste has no API, though several people immediately tried to make one.

Enterprise workflows entered the chat. Ramp CLI and Sierra Ghostwriter showed that agentic work was expanding beyond code into finance, customer experience, procurement, and operations. Once a workflow is represented as commands, permissions, and receipts, it becomes much easier to hand parts of it to an agent. It also becomes much easier to discover how much of the company was actually duct tape with a procurement policy.

Rate limits became social weather. Claude peak-hour limits and heavy internal dogfooding tweets made usage constraints part of the product conversation. That is what happens when customers stop treating a model as a feature and start treating it like electricity.

Representative signals

Date	Actor	Why it mattered	Link
Mar 16	OpenAI Developers	Codex subagents made parallel specialized work a first-party pattern.	View
Mar 16	Emil Kowalski	A design engineering skill turned blog knowledge into agent-operable taste scaffolding.	View
Mar 18	Vercel Developers	One Vercel plugin bundled 47-plus skills and subagents.	View
Mar 18	Stitch by Google	Stitch tied AI-native design canvas work to DESIGN.md and design systems.	View
Mar 23	Factory	Missions made long-running agents broadly available.	View
Mar 24	Figma	use_figma MCP made agents first-class participants on the canvas.	View
Mar 24	Claude	Auto mode reframed permissions as something an agent might negotiate on your behalf.	View
Mar 26	Ramp Labs	Ramp CLI turned finance operations into agent tools and skills.	View
Mar 28	Cheng Lou	Pretext reminded everyone that foundational UI engineering still matters in the agent age.	View

Engagement ledger

Tweet	Engagement
Cheng Lou on Pretext	65.5K likes, 23.8M views
Stitch AI-native canvas	41.7K likes, 18.6M views
Claude auto mode	39.4K likes, 7.7M views
Boris Cherny hidden Claude Code features	23.2K likes, 3.9M views
Figma use_figma MCP	10.4K likes, 17.8M views
Thariq on Figma MCP	6.3K likes, 1.1M views
Claude auto-fix in the cloud	6.1K likes, 1.5M views
Ramp CLI	2.5K likes, 596K views

The useful summary is that skills are becoming a kind of portable institutional memory. They carry the things a team would otherwise repeat in meetings: how to design, how to test, how to ask for permission, how to use the finance system without becoming a spreadsheet archaeologist. That is a less glamorous story than “the agent can do everything,” but it is more durable, because everything only becomes usable after someone teaches it the local customs.