Hermes just got 10x better...
ELI5/TLDR
Hermes Agent — a paid AI agent product that runs in Telegram — pushed an update with eight new features. The headline ones are persistent session memory (ask it what you did three weeks ago), background tasks that run while you keep chatting, computer control, native video generation, and an auto-kanban that breaks a goal into subtasks and farms them to sub-agents. Most of these aren’t new ideas, just newly bundled into one product.
The Full Story
Session recall, without burning tokens
The first feature is what Finn calls “session recall.” Ask Hermes what you worked on last Thursday, or two and a half months ago on a specific day, and it can tell you. The claim is that this runs programmatically — no token spend — rather than dumping the whole history into the context window every time.
“It’s just able to do it programmatically.”
For anyone who’s hit the memory wall with ChatGPT or Claude, this is the most genuinely useful item on the list. The rest is mostly orchestration sugar.
Background tasks
Type /background <task> and the agent goes off to work on it while you keep chatting. Finn fires off three at once — research AI startups, audit his last 50 newsletters, find trending AI videos — then asks about a basketball game and gets an answer mid-flight. The agent does the work serially in the background but doesn’t lock up the chat. Think of it like spinning up a coworker who knows not to interrupt you.
Grok 4.3 integration and X search
Two updates squashed into one: you can now plug your existing Grok (xAI) subscription into Hermes as a provider, and once it’s plugged in, Hermes can search live X/Twitter posts through it. Useful if you do content research and don’t want to pay twice for the same model. Less useful if you don’t already have a Grok login.
Codex CLI as a worker
This is the one a web dev should care about. Hermes can now spin up OpenAI’s Codex CLI as a worker process — even if your orchestrator is Claude Opus or GPT-5.5. So your expensive orchestrator handles planning and the cheaper Codex subscription does the actual code-writing in a dedicated folder. Finn combines this with background tasks and tells it to build a three.js first-person shooter in a single HTML file while he does other work.
The practical upshot: if you already pay for ChatGPT, you stop paying Claude rates to write boilerplate.
Computer use
The agent can now see your screen and click around. Finn demos it by asking Hermes to read his Notion calendar and add a 7pm event. It reads the calendar correctly, notices a conflict with Google I/O, but creates the event at 6:45 instead of 7. So — works, but not yet trustworthy for anything time-sensitive.
The interesting use case Finn floats: you’re out, your laptop is still running at home, you text Hermes from your phone and it does something on your desktop for you.
Native video generation
Plug in your Grok auth and the agent can generate videos directly inside the Telegram chat using Grok Imagine. Text-to-video or photo-to-video. The demo is a red dragon fighting a white horse, which Finn acknowledges is not exactly load-bearing for anyone’s workflow:
“Are you gonna ever need videos of dragons versus horses? No. But the fact that you can generate these videos natively without having to go to other websites is incredible.”
Auto-kanban
The last feature is the one Finn seems most attached to. Open the Hermes dashboard, click kanban, drop a goal into the “triage” column. Hermes breaks it into subtasks, moves them to “to-do,” and assigns each one to a different sub-agent.
His morning routine: write the day’s to-do list, paste it into triage, go make coffee. Come back to forty subtasks already in flight. Whether that ends with forty things actually done well is left as an exercise for the viewer.
Key Takeaways
- Hermes Agent lives in Telegram and uses your existing model subscriptions (Claude, ChatGPT, Grok) rather than charging for tokens directly.
- “Session recall” indexes past sessions programmatically so memory lookups don’t burn context — the most genuinely novel item on the list.
/background <task>queues work without locking the chat; you can pile up several and still talk to the agent.- Codex CLI as a worker lets a Claude/Opus orchestrator offload coding to a cheaper ChatGPT subscription — useful pattern for cost control if you already pay for both.
- Computer use works but isn’t precise — the calendar demo created the event 15 minutes off.
- The auto-kanban pattern (goal → subtasks → sub-agents) is the same pattern as Claude Code’s task tool or any orchestrator/worker setup, just exposed through a board UI.
- Grok integration is a soft up-sell: most features that involve “real-time” anything (X search, video gen) require a Grok subscription on top.
Claude’s Take
This is an update video for a paid product, made by someone who runs a paid bootcamp on that product. Calibrate accordingly. The walkthrough is competent, the demos mostly work, and Finn is upfront when something glitches (the calendar event lands at the wrong time). But the framing — “puts it ahead of OpenClaw,” “10x better,” “24/7 autonomous employee” — is the standard agent-influencer register. None of this is 10x anything.
That said, the actual feature set is a reasonable snapshot of where agent tooling is in mid-2026: persistent memory that doesn’t blow the context window, multi-task queuing, model routing across providers (Claude for orchestration, Codex for coding, Grok for X/video), and computer use. Hermes didn’t invent any of these. It’s just one of the products bundling them into one Telegram chat, which is genuinely a nicer interface than juggling four tabs.
The one feature worth borrowing as a mental model even if you never touch Hermes: the orchestrator/worker pattern with cheaper sub-agents doing the grunt work. That’s how Claude Code already structures sub-agents, and it’s how anyone building serious automation should be thinking — not “which model is best” but “which model for which job.”
Score 4/10. Useful as a market scan of what’s table stakes for agents right now. Light on substance otherwise.