Yesterday I just finished writing about Lobster 2.0, and today Hermes released Pantheon: half a year after defecting, it was worth it. Yesterday I had just published an article looking back at Little Lobster 2.0, in which, as "a veteran player who defected to Hermes," I talked a bit about the path my old friend has taken over the past six months. At the end of the article I left things hanging, saying I'd talk about my new friend another day. Then in the early hours of this morning, my new friend handed over the answer sheet itself. On September 1, Her

💡 What You Will Learn

Yesterday I just finished writing about Lobster 2.0, and today Hermes released Pantheon: half a year after defecting, it was worth it. Yesterday I had just published an article looking back at Little

📜 Table of Contents

• What's Actually Divine About the Pantheon (Bot Mode)
• What I Actually Felt Was Those "Old Pals" Finally Getting Promoted
• From v0.15 to v0.21, My Own Usage Has Changed Too
• Looking Back at Little Lobster, Was Half a Year of Defection Worth It

Just Finished Writing About Lobster 2.0 Yesterday, Today Hermes Dropped the Pantheon: Half a Year of Defection, Worth It

Yesterday I just published an article looking back at Little Lobster 2.0, chatting as a "veteran player who defected to Hermes" about the path my old friend has taken over the past six months. At the end of that article I teased that I'd talk about my new friend another day.

Then in the early hours of this morning, the new friend handed in its answer sheet on its own.

On September 1st, Hermes Agent officially released v0.21.0 on GitHub, codenamed "Pantheon." The official numbers: since the last major version, roughly 5,800 code commits, 2,475 merged PRs, 2,100 closed issues, and over 760 contributors. The previous version v0.20 was themed "Messenger God" — Hermes learned to relay messages to other agents; this version's theme is "Gods Assemble."

What's Actually Divine About the Pantheon (Bot Mode)

The headline feature this time is called Bot Mode, and it's built directly into the desktop client, enabled by default.

In plain terms: every AI agent under your command now has its own name, its own face (a system-generated deterministic avatar, which you can randomize or lock), and its own persistent chat window. You can create a Discord-style group chat, pull several bots into the same room, and @ whoever you want in the input box — that one comes out to work. Groups can be named and given avatars too.

The official wording puts it bluntly: in the past, "multi-agent" meant a bunch of plumbing and wiring; now it looks like a chat app full of coworkers.

There's also a companion hermes peer command: any of your agents can DM another agent by "account name," across configs and across gateways. For example, have a research bot hand what it found directly to a coding bot — and the replies land in their respective persistent chats, fully on record, not vanishing after the conversation ends.

What I Actually Felt Was Those "Old Pals" Finally Getting Promoted

Group chats and avatars look lively, but honestly, what really moved me as a long-time user were a few other updates — because I've been using the corresponding capabilities in scattered, makeshift ways for over half a year.

First: scheduled tasks have memory now. I run quite a few cron jobs with Hermes: grabbing an AI news radar every morning, monitoring web page changes, periodically summarizing data. The old scheduled tasks were like goldfish — every run was amnesiac, and the same news could get reported to you three times. This version hooks cron tasks up to persistent memory and a "continuity" toggle — results from this run carry into the next, and monitoring tasks that detect no changes even skip the model call entirely, saving tokens.

Second: sub-agents can be directed mid-flight. When doing heavy work I often dispatch several sub-agents to run in parallel at once, and previously once they were sent out, all you could do was wait. Now you can watch their status in real time, change direction mid-course, or just call a halt while preserving completed portions. The default parallel cap is now 10.

Third: MCP has a unified command center — one page to manage all tool servers, health checks that remind you to re-authorize before a tool call fails, and visibility into how many tokens each server has burned over the past 30 days.

Fourth: agents can finally drive the desktop client's built-in browser directly — navigate, click, read pages, all on their own. No longer just a window you can only "look" through.

None of these features individually count as explosive news. But together they point to one thing: back in the v0.15 days I was using a "very capable tool," and now it feels more like a team.

From v0.15 to v0.21, My Own Usage Has Changed Too

I started using Hermes in the v0.15 era, when the best part was "one assistant that can do everything": write scripts, make videos, build spreadsheets, look up information, run commands.

Now my daily routine looks completely different: in the morning, scheduled tasks have the news radar organized and waiting for me; before writing an article, I dispatch two sub-agents to separately verify data and dig through past drafts; when I need to operate external services, MCP gets connected on demand; writing cover images, editing videos, pushing WeChat official account drafts — all strung together in a single conversation. I always tell friends who want to try AI agents: stop fumbling around window by window, find a framework that lets you "hire employees" and delegate the work.

The Pantheon version essentially takes all the wild hacks I, as a heavy user, figured out on my own and conscripts them into a proper army.

Looking Back at Little Lobster, Was Half a Year of Defection Worth It

Yesterday I wrote a summary of Little Lobster 2.0, and my conclusion was: my old friend held it in for seven weeks to ship one major version, and it's proactively buckling its seatbelt — the direction is sound.

Today's answer sheet from Hermes tells a different story: it doesn't buckle up, it just floors it, and 5,800 commits drag the "AI assistant" species forward a full generation by force — from tool, to coworker, to OPC (a company you start yourself).

I still have both projects installed on my machine, and I haven't uninstalled either. But if you ask me where my main force is deployed, the answer is the same as half a year ago, and after today even more certain.

In 2026, is the AI in your hands still a "chat box," or do you already have a group of coworkers you can call by name, assign tasks to, and hand off work between? Let's talk in the comments.