Hermes Agent For Customer Support + Which AI Models To Pin To Triage, Drafts And Hard Cases: Real GoldieBench Scores

Hermes Agent for customer support doesn't need one model for everything: Hermes lets you pick a model per scheduled job, per chat channel and per Bot, so triage, drafting and hard cases can each run on a different one.

Below are the GoldieBench averages I read on the live models page on 7 November 2026, matched to the support jobs where Hermes lets you choose.

The examples are a dental clinic, an online shop and a design agency, and each pick says what the model does in that job.

Want my team to map models to your own support flow? Book a free SEO strategy session.

What GoldieBench Measures, And What It Doesn't

The GoldieBench methodology gives every model the same fixed prompt, one attempt, and the output must be one self-contained HTML file.

Julian scores each build from 0 to 10, and the models page ranks by the average.

So the numbers measure one-shot building: games, simulations and pages.

They don't measure how well a model writes a refund reply, reads a long email thread or follows a tone guide.

Treat them as a rough signal of overall strength, and test the shortlisted models on your own tickets before you commit.

A simple test: take ten past tickets with known good answers, run them through two shortlisted models in your drafting channel, and have the person who normally replies pick the better draft each time.

That tells you more about support writing than any build score can.

Four Places Hermes Lets You Choose A Model For Support

• Each cron job: the cron docs resolve a job's model as per-job pin first, then cron.model, then the main model. Pin with hermes cron edit <job_id> --model … --provider ….
• Each chat channel: channel_overrides in the messaging gateway set a model and a system prompt per channel or thread from one gateway.
• Each Bot: Bot Mode gives every Bot its own role, model, memory and skills.
• One-off hard prompts: the slash-command reference lists /moa <prompt>, which runs one prompt through the default Mixture of Agents preset and then restores your model.

That means a support desk can run a lighter model on the sorting job and a stronger one where words reach a customer, without two installs.

The AI Profit Boardroom trainings walk through these pins inside the Agent OS.

The Shortlist, With Live Scores

All figures are from the GoldieBench models page and the linked model pages as read on 7 November 2026.

The gap between first and seventh is half a point; the gap down to Gemini 3.6 Flash is bigger.

Triage Runs: A Mid-Table Model Is Enough

Sorting mail into five piles is a short, repetitive task with a human reading the result.

That's where I'd pin a model from the middle of the table, such as MiniMax M3 at 7.97 or Kimi K3 at 7.89, to the triage job only.

GoldieBench's Kimi K3 page describes it as tuned for long-horizon agent work with a context window of about one million tokens, which helps when a job reads a whole day of mail.

Pin it with hermes cron edit <job_id> --model … --provider …, and the rest of your setup stays on the main model.

The job runs the bundled himalaya and email-inbox-triage skills from the skills catalog.

The cron docs also let a job pin its own reasoning effort, from none up to ultra, with hermes cron edit <job_id> --reasoning-effort low.

Sorting rarely needs deep thinking, so a low setting on the triage job and a higher one on the weekly report is a sensible split.

Reply Drafts: Put A Top-Tier Model Where Words Reach Customers

Drafts are the part a customer eventually reads, so this is where the top of the table earns its place.

Fusion leads at 8.59 with 21 gold medals across 47 tasks, and Claude Opus 5 follows at 8.27 across 50.

Set the stronger model on your drafting channel with channel_overrides, so every draft written there uses it while the rest of your gateway doesn't.

A person still reads and sends every draft, because Hermes approvals cover risky commands and not outgoing messages, whichever model writes them.

Hard Cases: /moa And The Hermes MoA Score

Some tickets deserve more than one model's opinion: an angry complaint, a disputed charge, a legal-sounding letter.

GoldieBench lists Hermes MoA at 8.17 across 47 tasks, and its page describes a panel of frontier models whose drafts a chair model merges into one answer.

In Hermes, /moa <prompt> runs a single prompt through your default Mixture of Agents preset, then switches back.

Use it on the hard ticket only, keep the output as a draft, and have the owner send it.

Routing Group Chats Without Wasting A Big Model

If your support team runs as a group of Bots, Hermes can pick who answers a message that names nobody.

The Bot Mode docs say that choice is one short model call, and you can route it to a lighter model with auxiliary.group_router.

That keeps your top model for the drafts and away from simple routing decisions.

Bots also escalate judgment calls to a human with @user, which shows a "needs you" badge on the room.

Three Support Setups, Model By Model

A dental clinic

1. The triage job runs on Kimi K3 every morning and sorts appointment, insurance and emergency questions.
2. Emergencies go straight to the practice manager's Telegram.
3. Receptionists draft replies in a Slack channel pinned to Claude Opus 5, then send them from the clinic inbox.

An online shop on Freshdesk

1. The community freshdesk plugin reads new tickets on a schedule pinned to MiniMax M3.
2. Suggested answers are written as private notes on the main model, Fusion, and each note waits for approval because it's a write tool.
3. Staff send the public reply from Freshdesk.

A design agency in Slack

1. Client questions come into an internal channel set to GPT-5.6 Sol, which scored 8.16 across 50 tasks.
2. Disputed invoices go through /moa for a second view.
3. The account manager edits and posts every answer personally.

Safety Doesn't Change With The Model

Swapping in a lighter or stronger model changes nothing about what Hermes is allowed to do.

Dangerous-command rules for cron and webhook sessions live in your config, per the security docs.

So does trust: untrusted for an MCP server like the Intercom entry, which makes every write-capable call ask first.

Members of the AI Profit Boardroom share the exact approval settings they keep on support profiles.

For more on the hosted alternative, see our Grok Bot for customer support model guide.

Also On Our Network

• Agent Operatorsthe 8-step draft-only help desk setup, config blocks included
• AI Income Desksix small-business support jobs and the risks an owner should weigh
• AI Tool Verdictten support jobs scored out of 10, with the reason for each
• agentos.guidefive copy-paste draft-only prompts for a support desk in your Agent OS
• aiprofitboardroom.comthe complete Hermes Agent for Customer Support guide on the AI Profit Boardroom blog

Pin one model to your triage job today, and book a free strategy session if you want help choosing models for Hermes Agent for customer support.