Jonathan Sher 👨‍💻👾

tesla | ucla ai
building agents

[Jonathan Sher 👨‍💻👾 @officialjsher]

The wall isn't intelligence.

It's what survives a crash at 3am.

state → keys → restart.

https://t.co/ASTj50WxlZ

[Jonathan Sher 👨‍💻👾 @officialjsher]

https://t.co/RZD9UjvWHf

[Luke The Dev @iamlukethedev]

Hermes Bot Mode Tutorial

Bot Mode turns Hermes profiles into a roster of named, persistent teammates

Each Bot can have its own:

• Role
• Model
• Memory
• Skills
• Credentials
• Chat history
• Avatar
• Recurring routines

Here is how it works 🧵

1. Bot Mode is already inside Hermes Desktop

No extra installation is required

Open Hermes Desktop and select the Bots tab beside Sessions.

Important detail:

A Bot is still a normal Hermes profile underneath

Bot Mode simply gives profiles a better interface for managing them like a team

2. Create your first Bot

Click New Agent

The quick setup asks for:

• Name
• Title
• Description

Open Advanced when you want more control.

You can:

• Start fresh
• Clone another Bot
• Create an empty profile
• Pin a specific model and provider
• Write a custom SOUL.md
• Select its skills
• Select its toolsets
• Select its MCP servers

Different Bots can use completely different models and capabilities

3. Every Bot gets its own permanent chat

The moment you create a Bot, Hermes creates and pins its main Bot Chat.

Clicking that Bot always takes you back to the same conversation.

Inside this permanent chat, /new and /reset become /compact

That gives the Bot fresh working context without breaking the ongoing relationship or creating a disconnected session

4. Give every Bot a specific job

A good Bot should not try to do everything.

Example team:

Researcher

Finds sources, compares options, and prepares reports.

Developer

Reads repositories, writes code, and runs tests.

Reviewer

Reviews pull requests, checks security, and looks for regressions.

Operations

Monitors systems, runs scheduled checks, and reports incidents.

Support

Reads requests, classifies issues, and prepares responses.

Each Bot gets the memory, instructions, model, and tools needed for its role.

5. Add recurring routines

A Bot can own recurring tasks directly from the Routines pane.

Examples:

• Summarize my inbox every morning
• Check deployments every hour
• Review open Jira tickets every weekday
• Send me a daily cost report
• Scan failed jobs every 15 minutes

These routines are normal Hermes cron jobs

They also appear in hermes cron list, and every result lands inside that Bot’s own chat history.

6. Hand work from one Bot to another

From any Bot chat, mention another Bot

Example:

researcher compare these three API providers

The active Bot hands off the request, waits for the response, and brings the answer back with attribution

This lets one Bot remain responsible for the main task while calling specialists when needed.

7. Create group chats for your Bots

Organize Bots into a group, then open the group chat.

A group can contain between 2 and 6 Bots

Hermes gives them up to three rounds to discuss your message

You can mention specific Bots, or let every relevant Bot decide whether it has something useful to add

Bots can also call another Bot into the discussion or escalate a decision back to you with user

Hard limits prevent the room from talking forever

8. Your Bots can live on different machines

Register your local computer, remote gateways, SSH hosts, or Hermes Cloud instances under Connections

The Bots from every machine appear in one roster.

A researcher can run in the cloud

A developer can run on your workstation

An operations Bot can live on a VPS

They can still message each other and participate in the same group chat while their work remains on the machine that owns each profile

9. Bot Mode still works with the CLI

Because every Bot is a Hermes profile, the same agent is available outside Desktop.

Chat with one:

hermes -p researcher chat

Its files, memory, and skills live under:

~/.hermes/profiles/researcher/

Its routines appear in:

hermes cron list

Desktop is the visual layer
Profiles remain the foundation

[Jonathan Sher 👨‍💻👾 @officialjsher]

@witcheer the contributor wall does the talking

[Jonathan Sher 👨‍💻👾 @officialjsher]

skills only stick when the file auto-invokes, otherwise it's another doc nobody opens

[nyk @nykdotdev]

@kirillk_web3 The Chinese dev bit lands because workflow patterns rarely fit cleanly in docs.

Claude Code + SKILL.md: build your first auto-invoking AI workflow in 15 minutes
https://t.co/yUxlhxft9E

[Jonathan Sher 👨‍💻👾 @officialjsher]

@neil_xbt @threejs does the 120 fps hold once a second player is in the scene, or is that still single client

[Jonathan Sher 👨‍💻👾 @officialjsher]

@yuki_buildsAI shipped a piece tonight about exactly this, left the tab unattended and the run looked green the whole 5 days until i sat and watched one

[Jonathan Sher 👨‍💻👾 @officialjsher]

@elshayib_ the 400k ceiling is the useful part, not the 1M headline

[Jonathan Sher 👨‍💻👾 @officialjsher]

Chat is the debugger.
The loop is the product.

[Jonathan Sher 👨‍💻👾 @officialjsher]

https://t.co/xMmkCLYHHe

[Jonathan Sher 👨‍💻👾 @officialjsher]

https://t.co/xMmkCLYHHe

[Machina @EXM7777]

if i were you... i'd spend a day or two building a simple Hermes setup right now

this is probably the highest leverage play you can make with AI agents

i don't think any other project is as impactful for your work

in just a few days they shipped "bot mode", which is basically an UI improvement for ppl who loved Grok Bot

now we have 900k context using GPT models (what works best with Hermes for me)

every single week your agent is improving via updates only, this is awesome

[Jonathan Sher 👨‍💻👾 @officialjsher]

@mrfundman Yes plz

[Jonathan Sher 👨‍💻👾 @officialjsher]

the ranking is right but rung 4 keeps showing up because most agent work isn't code. nothing lints "did it update the right row". so cheaper review, not less review.

[lauren @poteto]

every time you intervene and correct your agent, you should think about how to eliminate it entirely.

in order of value:
1. categorically eliminate the problem through better architecture or choice of data structures
2. turn it into a lint rule or test so CI catches it
3. turn it into a skill or rule
4. have humans review the code to catch it (ngmi)

[Jonathan Sher 👨‍💻👾 @officialjsher]

@phibrowser @poteto cheaper interventions, yeah. That's the whole game outside a repo.

[Jonathan Sher 👨‍💻👾 @officialjsher]

@Teknium 2,536 and the review queue is the actual bottleneck now

[Jonathan Sher 👨‍💻👾 @officialjsher]

@poteto where does step 2 live for you, CI or a pre-commit hook? the ones I put in CI got noticed way too late

[Jonathan Sher 👨‍💻👾 @officialjsher]

@ziwenxu_ the token math always wins eventually

[Jonathan Sher 👨‍💻👾 @officialjsher]

31,650 to 3,800 requests is the reminder that your agent's economics live in someone else's pricing page. If a sub is the whole plan, the default model changing under you is the outage.

[Ziwen @ziwenxu_]

OpenCode just got nerfed by 10x.

DeepSeek raised their prices, so the sub moved with them. 31,650 requests per 5 hours down to 3,800.

GPT-5.6 Luna now gets more usage on that same $10 sub than DeepSeek V4 Flash does, which is backwards from the whole reason most of us bought it.

So we might as well run Luna in there instead. It's the same sub and the same router, we just change which one is the default.

If you bought OpenCode Go for the DeepSeek volume then it's time to switch into Luna.

[Jonathan Sher 👨‍💻👾 @officialjsher]

@nykdotdev Right, though the checkpoint has to land before the side effect or you're just guessing. Ordering problem more than a retry problem.

[Jonathan Sher 👨‍💻👾 @officialjsher]

https://t.co/bP6sFBGDKp 272K locked to 900K is the part that matters more than the 1M headline. the ceiling most of us kept hitting was the sub tier, not the model

[Teknium 🪽 @Teknium]

Upped ChatGPT / Codex Subscription users limits to 900K in Hermes Agent, enjoy!

[Jonathan Sher 👨‍💻👾 @officialjsher]

@viemccoy blind or it's just vibes

[Jonathan Sher 👨‍💻👾 @officialjsher]

@BlendiByl timestamps help, sure. but then the model starts doing date math in its head and gets that wrong too. curious if you'd stamp every message or just the last few