Six vendors skipped the inventor

Hey there 👋

For about a year, every agent platform pitched you the same deal: pick our runtime, live inside our walls, and life gets simpler. I never loved that trade, and this week the wall grew a door. Six of the biggest vendors agreed on one plugin format that installs across all of their clients. Meta put a 30-billion-parameter agent model on Hugging Face for anyone to download. The building blocks of an agent are walking out from behind the walls.

So here is my bias, plain: bet on the pieces you can carry out the door, and keep your money in the one layer a vendor can truly lock. Today is about that turn, and about what you can package or download because of it.

Ron

The Big Thing

Six vendors, one plugin. The part they left out is the story.

For a year the coding-agent world made you choose a home. Build an extension for one assistant and it stayed there, useless everywhere else. On August 6 that changed. OpenAI, AWS, Microsoft, Vercel and Cursor's maker Anysphere, with GitHub in the mix, shipped Agent Plugins v1.0.0, a vendor-neutral format for packaging an agent extension once and running it in many places.

The format is refreshingly boring, which is the point. A plugin is a folder. At its root sits a plugin.json manifest. Drop your reusable skills in a fixed skills/ folder. Add an optional mcp.json if the plugin ships tool servers. There's a namespaced escape hatch for client-specific behavior, so one vendor's quirk doesn't break the shared contract. That's the whole spec.

The payoff for a builder is direct. Package the extension one time and it loads in ChatGPT, Codex, GitHub Copilot, VS Code, Cursor and AWS's Kiro. Six clients, one build. Vercel kicked off the proposal, and the founding steering group is AWS, Anysphere, Microsoft, OpenAI and Vercel. If you maintain an internal tool that only your team's assistant of choice can see today, this is the week that stops being a lock-in problem and starts being a copy-the-folder problem.

Now the part they buried, and the reason this is today's lead. The format is built on two primitives: Agent Skills and MCP. Anthropic invented both. And Anthropic is not on the steering committee. Read that again. Six rivals standardized a distribution format around a competitor's two inventions, then left the inventor off the governance table.

Grade the launch accordingly. A v1.0.0 spec with six logos on it is an agreement, not proof. Nobody has shipped a year of real multi-client parity yet, and specs drift the moment one client's roadmap pulls hard. The governance gap is the sharper risk: when the company that owns the underlying primitives has no seat, the standard can fork the day interests diverge, and you're the one holding plugins built to a contract that split in two.

Why it matters: distribution just got cheap. If your extension is portable, your leverage stops being chained to any single assistant's store.

Ship it? Yes, with your eyes open. If you already run MCP servers or Agent Skills, repackage them in the new format now and take the free reach across six clients. Then watch the governance question like a hawk, because a standard missing its primitives' inventor is one disagreement away from a split.

Sources: OpenAI dev blog, AWS OSS blog, Vercel

Tour de Headlines

🦙 Meta open-weighted a coding agent that runs on one GPU. Meta Superintelligence Labs, under Alexandr Wang, put Muse Glimmer on Hugging Face: 30B parameters, Apache-2.0, distilled from the bigger Muse Spark. Quantized to 4-bit it fits under 20GB, so a single 24 or 32GB consumer card runs it. It pairs with Muse Code, a terminal coding agent. That gives you a fourth serious coding-agent vendor next to Anthropic, OpenAI and Google, plus a self-host hedge you own outright. One flag on the scoreboards: the category-best marks against Gemma4 and Qwen3.6 come from Meta's own harness and charts, so read them as vendor framing. And the flagship Spark weights are still "coming soon," so today you download the smaller distill. MarkTechPost

🛡️ Three agent-security rounds landed in one week, $270M total. They rhyme with the same pitch: inspect what an agent is about to do and block the action before it runs, instead of scanning the prompt after the fact. Zenity raised a $125M Series C led by Norwest. Obsidian Security took $85M at a $1.1B valuation. Oligo pulled $60M for catching exploits at runtime. We've covered Zenity and Obsidian before, so what's fresh is the cluster: three raises, one week, all aimed at the runtime-action layer. Funding is a demand signal. It measures appetite and nothing more, and right now that appetite points hard at governing what agents can touch. StartupHub

📊 Open-weight models are within about 7 points of the frontier on real tool use. Toolathlon, from HKUST NLP and headed to ICLR 2026, scores general tool use across 600+ real tools and 32 apps like Notion, Kubernetes and BigQuery over long, roughly-20-turn tasks. A public snapshot puts Claude Opus 5 on top at 80.6%, with Kimi K3 at 73.2% and Qwen3.8 Max at 72.5%. One honest asterisk on that 80.6: it's Anthropic's own-harness Pass@1, and the published leaderboard reports a different figure because it handles null attempts another way. So take the gap as directional. The builder takeaway holds either way: an open-weight model may already clear the bar your agent needs, so right-size the model instead of reaching for the priciest one by reflex. GitHub

Tool of the Day

🧩 Claude Managed Agents: Skills straight from a GitHub repo

What it's for: shipping a team's reusable agent skills from version control, with no manual upload step.

Point a Claude Managed Agents session at a GitHub repo as a resource, and any Skills sitting in its root .claude/skills folder get auto-discovered at session start and handed to the agent. Your skills live in git, get reviewed in pull requests, and reach the agent the moment a session opens. There's a nice rhyme here: the one company left off the Agent Plugins standard is shipping its own clean path to distribute Skills, the exact primitive that standard packages. One catch to plan around: the scan runs once, at session start. Commits you push mid-session won't show up until you open a fresh one.

Read the docs

Worth a Click

• 🎙️ OpenAI put voice control of your agents on the desktop. ChatGPT for macOS and Windows now lets you drive your computer and steer several Codex or ChatGPT Work agents by voice, running on GPT-Live's full-duplex audio. Treat it as a direction signal while the rollout scope settles. VentureBeat
• 🧰 Meta Muse Code is the one-command way to run Glimmer this afternoon. The terminal agent spins up parallel sub-agents in isolated git worktrees with a crash-safe event log on macOS and Linux, though its 12x-cheaper contributor tier trades the discount for rights to train on your code. TechCrunch

Delight

✨ The slide that priced a billion coworkers

Zenity raised its $125M to secure "the era of 1 billion AI agents." Sit with that number for a second. The pitch deck quietly swapped seats for bots, and whatever you make of the count, it's the unit enterprise budgets are starting to underwrite. Take it as a marketing frame stapled to a funding round. Fun to notice, easy to over-read.

Source: Zenity

The Bottom Take

Own the portable building blocks. Rent the runtime.

For three issues straight I've argued the moat is hardening into the runtime, the harness and the governance layer you run agents on. That's still true. Today pulls the other way at the same time, and both hands are right. The building blocks are going open and portable in a hurry: one plugin format across six clients, a 30B agent model you can download and keep, open weights within a few points of the frontier on real tool use. The stuff you build with is stepping out from behind the walls.

So stop treating this as a single bet. It's two. Put your tooling on the open, portable formats and weights, the pieces you can carry to any vendor and never rebuild. Package your extensions to the shared plugin spec. Keep a downloadable open-weight model in your back pocket for the jobs where it clears the bar, and pay frontier prices only where the task earns it.

Then spend your real leverage on the one layer nobody can hand you for free: the runtime you operate, the data and the judgment and the reliability that only run in your shop. That's where a vendor still owns you, and it's the only place worth being owned. Before next year, get the split straight: the commodity pieces belong on open formats, and your leverage belongs in the layer only you can run. Build portable, and keep your money where the walls still hold.

Ron