Three unrelated stories from Japanese tech outlet Publickey this week say a lot about where developer tooling is heading — one about making AI agents less siloed, one about a well-known open-source builder starting fresh, and one that’s just a nice reminder that technical explainers don’t have to be boring.

Agent Plugins 1.0.0: a shared language for AI agent configuration

The most consequential item is the announcement of “Agent Plugins 1.0.0,” a new specification backed by AWS, Microsoft, OpenAI, Anysphere, and Vercel, with Google also signaling support. According to Publickey’s report, the goal is to let “skills” and MCP (Model Context Protocol) server configurations be shared across different AI agents rather than being locked to a single vendor’s implementation.

This matters because the AI agent ecosystem has quietly recreated the exact fragmentation problem that plagued early cloud and DevOps tooling. Every major AI vendor — OpenAI, Anthropic, Google, Microsoft — has its own agent runtime, and each has been rolling out its own notion of “skills” (reusable capabilities an agent can invoke) and its own way of wiring up MCP servers, the increasingly standard protocol Anthropic introduced for connecting LLMs to external tools and data sources. MCP itself solved one layer of the problem: it gave agents a common way to talk to tools. But it didn’t solve the layer above that — how you package, describe, and configure a set of skills and MCP connections so they behave consistently no matter which agent framework is running them.

Agent Plugins 1.0.0 is essentially trying to be that packaging layer. If it gains traction, a developer could define a bundle of skills and MCP server configs once and expect it to work whether the underlying agent is Claude, GPT-based, Gemini, or something running inside AWS’s or Microsoft’s agent platforms. That’s a meaningful bet against vendor lock-in, and the fact that direct competitors like OpenAI, Microsoft, AWS, and Google are all lining up behind the same spec (even if with varying levels of formal commitment) suggests there’s shared recognition that fragmentation is bad for adoption across the board. It’s the same pattern we’ve seen before with container runtimes, Kubernetes CRDs, and OpenTelemetry — once enough vendors decide interoperability grows the pie faster than lock-in protects their slice, a standard emerges surprisingly quickly.

The real test, as always, will be whether “1.0.0” actually means stable and implementable, or whether it’s an early draft that changes shape once real-world agent builders start using it in production.

Mitchell Hashimoto’s next act: Superlogical

Separately, Mitchell Hashimoto — co-founder of HashiCorp and, more recently, the independent developer behind the terminal emulator Ghostty — has announced a new company called Superlogical. Per his own announcement quoted in Publickey’s coverage, the company will start by building a terminal multiplexer, with a stated ambition that “the entire vision is much larger” than that.

A terminal multiplexer might sound like a narrow, almost nostalgic project in 2026 — tmux has been the default answer for decades, and Zellij has picked up a modern, Rust-based following in recent years. But Hashimoto’s track record suggests he doesn’t pick small problems. HashiCorp’s tools (Terraform, Vault, Consul) succeeded by taking unglamorous infrastructure plumbing and making it dramatically more usable and composable. Ghostty followed the same playbook for terminal emulators: take a category everyone assumed was “solved” and rebuild it with modern engineering (GPU rendering, tight platform integration) and obsessive attention to developer experience.

Starting Superlogical with a multiplexer, while hinting at a “much larger” vision, reads like the same strategy again — use a well-understood, developer-beloved tool as the entry point to build trust and distribution before expanding into whatever the bigger platform play actually is. Given his history of building tools that quietly become infrastructure defaults, it’s worth watching closely, even before the full vision is public.

And for something lighter: the 2026 IT manga roundup

Rounding things out, Publickey published its annual summer collection of IT-themed manga — comics that explain technical concepts like how AI systems work or what data center operations actually look like, aimed at making dense subject matter more approachable. It’s a small but recurring feature (Publickey runs it every summer), and worth a look if you want technical explainers that don’t read like documentation. Given that two of this year’s other big stories are a dense interoperability spec and a not-yet-explained “vision,” a comic-format primer on AI internals might be the more digestible read of the week.