Claw chronicles: When 'everywhere' means 'nowhere specific'
Hermes Agent v0.9 dropped four days ago and the release notes read like a feature wishlist someone accidentally committed. 487 commits. 269 merged PRs. 16 messaging platforms. A web dashboard. Android support. Fast Mode. Background process monitoring. They called it “The Everywhere Release.”
Nous Research went from first release to 27,000 GitHub stars in two months. That’s a growth curve that would make a crypto startup jealous. I say that as someone running a NanoClaw instance that supports exactly the messaging platforms I’ve explicitly configured. Reading through the Hermes v0.9 changelog made me feel like I was watching someone install a second kitchen in a house that already has a perfectly good kitchen.
The three answers
Three different projects just gave three completely different answers to “where should your AI agent live?”
Hermes says: everywhere. Sixteen platforms. WhatsApp, Telegram, Discord, Slack, iMessage, WeChat, LINE, Matrix, Signal, and on and on. The thesis is that your agent should be a persistent presence across every communication channel you use, with a unified brain behind all of them. Switch platforms, the agent follows.
Cursor 3 says: in your editor. They launched on April 2 and the Reddit post nailed it: “Cursor 3 just replaced the code editor with an agent management console.” Agent Tabs let you run multiple agents side by side. /worktree creates isolated git worktrees. /best-of-n runs the same task in parallel across different approaches. The editor has been demoted to dashboard for your coding agents.
Codex CLI says: in your terminal. OpenAI’s open-source Rust agent just hit v0.121 alpha. It lives in the shell, where developers already live. It reads your files, edits your code, runs your tests. No GUI, no web dashboard, no mobile app. Just codex and a prompt.
Three form factors. Three philosophies. And I think the tension between them tells us something important about where agents are actually headed.
The everywhere problem
“Everywhere” sounds great until you think about what it actually means in practice.
Hermes Agent supporting 16 platforms sounds incredible until you think about what that actually means in practice. Each platform has its own message format, its own attachment handling, its own threading model, its own rate limits, its own authentication quirks. iMessage requires a Mac running somewhere. WeChat has its own ecosystem entirely. Signal’s encryption model is fundamentally different from Telegram’s.
Supporting all of these means maintaining 16 different integrations, each with their own edge cases and failure modes. When something breaks — and it will — which platform gets fixed first? When a new platform API changes, how quickly does a project with 487 commits in a single release cycle respond?
I’ve been running NanoClaw long enough to know that even supporting two messaging platforms well is harder than it looks. There’s a reason NanoClaw’s codebase is 500 lines. Every platform you add is a surface for bugs, a maintenance burden, and a potential security issue.
The “everywhere” approach also creates a weird user experience problem. If I can talk to my agent on WhatsApp, Telegram, Discord, iMessage, and WeChat simultaneously, which conversation is the “real” one? How does context carry across? If I tell my agent something sensitive on Signal, does that context bleed into the Discord conversation? Hermes solves some of this with per-platform session management, but the fundamental tension remains: more surfaces means more state synchronization problems.
Why Cursor’s bet is the scariest
Of the three, Cursor 3’s approach is the one I find most thought-provoking and the one I think the industry is underestimating.
Cursor isn’t trying to put agents in your messaging apps. They’re making the agent the primary interface and relegating the code editor to a supporting role. When they say “the third era of AI software development,” they mean agents that tackle large tasks independently over extended periods with minimal human input. The IDE becomes a monitoring dashboard.
This is a fundamentally different mental model. Instead of “I ask my agent to do something and it does it,” it’s “I have a fleet of agents running in parallel and I check in on them.” That’s the jump from single-player to management sim.
The /best-of-n feature is the canary in the coal mine. It runs the same task multiple times in parallel and lets you pick the best result. That goes beyond acceleration. It’s a fundamental change in how you approach reliability. It implicitly acknowledges that AI outputs are probabilistic and the way to get reliable results is to generate multiple candidates and curate. Once you internalize that, the whole “one agent, one task” model starts to feel naive.
My concern with Cursor’s approach is lock-in. You’re not just locked into an editor. You’re locked into their agent runtime, their cloud infrastructure, their pricing model. At least with a terminal agent like Codex CLI, you can swap models and run everything locally. With Cursor, the agent is the product.
The terminal dogma
Which brings me to Codex CLI, and why I think the terminal crowd has the most intellectually consistent position.
The terminal is the least opinionated interface you can build an agent on. It doesn’t dictate how you organize your work. It doesn’t impose a visual paradigm. It doesn’t try to be your “AI home.” It’s just a pipe between you and the model, with your filesystem as the shared context.
Codex CLI’s v0.121 additions (stdin piping, improved Linux sandboxing, workflow resume fixes) are all infrastructure improvements. They’re not chasing platform count. They’re making the single experience of “coding agent in your terminal” better.
There’s something refreshingly honest about this. The terminal doesn’t pretend to be anything other than what it is. It doesn’t have a marketing-friendly name like “The Everywhere Release.” It’s just a tool that does a thing.
Where I land
I run a NanoClaw instance. I talk to it in group chats. It runs in a container. It can’t reach 16 platforms, it can’t open a web dashboard, and it can’t run background tasks across multiple git worktrees. What it can do is respond to me reliably in the places I’ve told it to be, with a codebase I can audit and a blast radius I can contain.
I think the claw ecosystem is about to hit the same wall that every “supports everything” platform hits: the realization that breadth and quality are inversely correlated. The project that supports 16 platforms will have more users. The project that supports two platforms really well will have more satisfied users.
Hermes v0.9 is an engineering achievement. 487 commits in one release cycle is impressive. But I’d trade every single one of those platform integrations for rock-solid reliability on the ones that matter most.
One prediction
By the end of 2026, at least one major claw project will announce a “platform diet,” dropping support for half their integrations to focus on making the core experience excellent. And their user satisfaction scores will go up, not down.
The “everywhere” era is fun. The “somewhere specific” era is where the real products get built.
Claw Chronicles is a daily dev diary about the AI agent ecosystem. I run NanoClaw and have opinions. Neither is neutral.