Hermes Agent Slack + Which AI Models To Run Behind The Bot, With Real GoldieBench Scores

Hermes Agent Slack works with whatever model your Hermes profile points at, so the model you choose decides how good every thread reply, channel answer and scheduled post will be.

The Slack side is the same whichever model you run. The quality isn't.

Below are the GoldieBench scores I read on the live pages on 11 November 2026, what they mean for a team chat bot, and which model I'd put behind which kind of Slack work.

If you'd rather talk the choice through with my team, book a free SEO strategy session.

What A Slack Bot Asks Of A Model

A Slack bot has three very different jobs, often on the same day.

• Quick answers to a teammate's @mention in a busy channel.
• Careful work in a client channel, where a weak draft costs trust.
• Unattended runs, like a cron job that posts a brief to the home channel before anyone is online.

The Hermes Slack docs cover the wiring: Socket Mode, the two tokens, the allowlist and the mention rules.

They don't pick a model for you. That's where the bench comes in.

For model picks across a whole firm rather than one chat app, see our Hermes Agent for business model guide.

The GoldieBench Scores I'd Shortlist

Cloud ranks come from the main GoldieBench leaderboard, which lists 33 models, 50 tasks and 1,310 live demos.

Fusion sits at #1 with 8.59, but it's an OpenRouter entry, so I've kept it off a list of single models to run behind one bot.

What The Scores Measure, And What They Don't

The methodology page is clear: one fixed prompt per task, one shot, one self-contained HTML file out.

Every score is Julian's own 0 to 10, averaged across the tasks a model ran.

So the bench reads first-try quality on build tasks.

It doesn't score Slack replies, thread etiquette or how well a model summarises a channel.

Use the gap between 8.27 and 7.08 as a quality signal for work where nobody gets a second try, not as a promise about your exact channel.

Client Channels: Pick The Top Single Model

An agency's client channel is where a sloppy reply hurts most.

Claude Opus 5 has the highest single-model average on the board at 8.27 across 50 tasks, with 13 gold medals on its page.

GPT-5.6 Sol sits right behind at 8.16, and its page says it was benched at medium reasoning effort.

The Sol page also lists an outside reference figure of 0.910 for prompt-injection defence. That number was pulled from an agentos.guide guide, not from the bench's own medals, but it's worth weighing for a bot that reads messages from many people.

Whichever you pick, keep "draft only" in that channel's prompt, so a person sends anything that reaches the client.

Busy Team Channels: A Fast Flash-Tier Model

An ops channel at a plumbing firm fires off lots of short questions: who's on call, where's the van, what's the job reference.

Gemini 3.6 Flash scores 7.08 across 50 tasks, and its page calls it faster and lighter on tokens.

The same page labels it a Flash tier, not a flagship, and suggests it for routing the everyday work while a flagship handles the hard part.

That's the right shape for a channel where speed beats polish.

Keep the full-detail status line off in that room too. Setting live_status to verb shows what the bot is doing without printing file paths or commands.

Overnight Cron Posts: Kimi K3

A Slack home channel often gets one long job a night, like a weekly audit posted for the Monday stand-up.

Kimi K3 scores 7.89 across 50 tasks with 8 golds.

Its page says it runs in the Agent OS as the kimi-k3 Hermes profile for long unattended agent runs.

The trade-off is listed plainly: it's slow on hard tasks.

For a post that only has to be ready by morning, that's fine. Point the job at the home channel with deliver: slack, as the Hermes cron docs describe.

High-Stakes Questions: Hermes MoA

Some questions deserve more than one model's opinion, like a partner asking for a risk summary before a client meeting.

Hermes MoA scores 8.17 across 47 tasks.

Its page describes a panel of frontier models drafting in parallel, with a chair model merging them into one answer, run from the Mixture tab in the Agent OS.

The page is honest about the cost in waiting: latency is the slowest draft plus the merge pass.

That's too slow for chat banter, and worth it for the rare question that matters.

Members of the AI Profit Boardroom get the Agent OS with the Mixture tab already in place.

Private Data: Check The Local Leaderboard

An accounting firm may want month-end questions answered without the text leaving the office.

Socket Mode already means the bot needs no public URL. A local model keeps the inference on the machine too.

On the local leaderboard, Qwable 5 27B Coder leads at 7.14 across 41 tasks, running through Apple MLX.

The next local entry drops to 4.83, so the gap after first place is wide.

Remember the bench is build tasks, so test a local model on a week of real channel questions before trusting it with month-end.

Hermes supports local setups through its local models guide.

Codex-Backed Peer Bots: Keep Effort At High Or Lower

The Slack page carries one model tip worth copying exactly.

For Codex-backed peer-agent channels, it recommends agent.reasoning_effort: high or lower.

At xhigh, a turn can spend everything on hidden reasoning and never post visible text.

Pair that with allow_bots: mentions so peer bots only answer when tagged, which the docs call the safest bot-to-bot mode.

Switching Models Without Leaving Slack

You don't have to pick one model forever.

The slash command reference says a plain /model switch lasts only for the session unless you add --global, and --once applies to the next turn only.

/model --once <top-model>        # one careful answer, then back
!model <fast-model>              # inside a Slack thread, use the ! prefix

The ! form matters because Slack itself refuses slash commands in thread replies.

The reference also warns that switching mid-conversation resets the prompt cache, so the next turn re-reads the whole conversation.

And since v0.23.0, one gateway owns each session, so a switch made in Slack shows in the same conversation everywhere.

Set a sensible default, keep the top scorer one /model --once away, and review a week of replies before you change anything.

Also On Our Network

• Agent Operatorsthe six-step Socket Mode setup and five config lines to add first
• AI Income Deskfour small-business Slack setups and the access rules owners must know
• AI Tool Verdicta nine-point scorecard of the Slack bot, with who should skip it
• agentos.guidefive copy-paste prompts that let your agent set up Slack itself

Pick a default from the scores above, keep a flagship a command away, and join the AI Profit Boardroom or book a free call to tune the model behind your Hermes Agent Slack bot.