Hermes Agent Hooks + Which AI Models To Pair With Them, With Real GoldieBench Scores For Opus 5, Hermes MoA, Kimi K3 And Gemini 3.6 Flash

Hermes agent hooks don't care which model you run, and that's exactly why the model choice and the hook setup should be planned together.

Hooks decide what reaches the model and what the model is allowed to do; the model decides how good the work is once it gets there.

Below are the GoldieBench scores I read on the live pages on 2 November 2026, what they mean for a hooked Hermes setup, and how the new pre_llm_call block changes the picture.

If you'd rather my team matched models and guardrails to your work, book a free SEO strategy session.

Where The Model Sits In A Hooked Turn

The Hermes hooks docs put pre_llm_call once per user turn, before the tool loop starts.

Its payload includes a model field, the model identifier for that session, alongside platform and the user's message.

Then the model runs, and every tool it calls passes through pre_tool_call first.

So hooks wrap the model on both sides: the input it receives and the actions it takes.

A higher-scoring model gets more right first time, while hooks catch the cases where any model would get it wrong.

The GoldieBench Scores For A Hermes Shortlist

Positions follow the order on the GoldieBench models page, which compares 33 frontier models.

Fusion leads that page at 8.59 across 47 tasks, but it's an OpenRouter multi-model panel, so I've kept it off a list of models you'd set as a Hermes session model.

What These Scores Mean For Hooks

GoldieBench gives every model the same fixed prompts, one shot each, scored 0 to 10 by Julian.

That's a read on first-try build quality, not on how well a model behaves under a policy.

No score here measures whether a model tries a destructive command or pastes a client ID back into a tool call.

That's the job hooks do, and they do it the same way for a 7.01 model as for an 8.27 one.

So use the scores to pick the model and the hooks to fence it in.

One more reason to keep them separate: a model swap needs no hook changes, because the rules live in your profile config, not in the prompt.

Switch from Sonnet 5 to Opus 5 tomorrow and every gate you wrote today still fires on the first tool call.

Top Scorers: Guard The Tools, Not The Talent

Claude Opus 5 posts 8.27 across 50 tasks with 13 gold medals on its page.

GPT-5.6 Sol sits at 8.16 across 50 tasks.

When I run a model this strong, the risk isn't bad output; it's confident action.

That's where the tool gate earns its place, starting with the docs' block on destructive terminal commands:

hooks:
  pre_tool_call:
    - matcher: "terminal"
      command: "~/.hermes/agent-hooks/block-rm-rf.sh"
      timeout: 5

Add fail_closed: true on any gate that scans for secrets, so a crashed scanner blocks the call instead of waving it through.

Hermes MoA: Watch The Panel Calls

Hermes MoA scores 8.17 across 47 tasks, and its GoldieBench page describes it as a panel of frontier models merged by a chair.

In the hooks catalog, MoA is listed among the auxiliary LLM calls, next to titling, compression, vision and approval.

Each attempt fires pre_auxiliary_call and post_auxiliary_call, which carry the model, provider, usage and duration.

Those are observer hooks, so they can log but not block.

If you run MoA, register one to see how many panel calls a turn really makes.

Kimi K3: Hooks For Unattended Runs

Kimi K3 scores 7.89 across 50 tasks with 8 gold medals.

Its GoldieBench page says it's wired into the Agent OS as the kimi-k3 Hermes profile and used for long unattended agent runs.

Unattended is the key word.

The gateway, cron and CI never show the first-use hook prompt, so a hook you haven't approved stays unregistered on exactly the runs that need it most.

Every demo on GoldieBench was built inside the Agent OS, which members of the AI Profit Boardroom can install for their own runs.

Approve hooks in the profile's allowlist before a long run, and check them with hermes hooks list.

Shell hook payloads include a profile field, so one script can apply the right rules to the kimi-k3 profile.

Gemini 3.6 Flash And Sonnet 5: Lighter Lanes

Gemini 3.6 Flash scores 7.08 across 50 tasks; Claude Sonnet 5 scores 7.01 across 47.

Those are the lanes I'd hand routine, high-volume chat to.

High volume makes context injection more valuable, because every turn starts from the same facts.

A pre_llm_call hook can add those facts to the user message each turn without touching the system prompt, which the docs say keeps the prompt cache intact.

POLICY = "Never execute commands that delete files without explicit user confirmation."

def guardrails(**kwargs):
    return {"context": POLICY}

def register(ctx):
    ctx.register_hook("pre_llm_call", guardrails)

That's the docs' guardrail example, and it works the same on any of these models.

The New Block: Zero Model Calls, Whatever You Run

Commit ad9229c843, merged to main on 1 Nov 2026, lets pre_llm_call return {"action": "block", "message": "..."}.

The turn then makes zero model calls, so a blocked prompt never reaches Opus 5, Kimi K3 or any other model on the shortlist.

A 1 Nov follow-up stores the prompt only as [Message withheld: a pre_llm_call policy hook blocked it.], so later turns and compaction never forward it either.

Another fix stopped a blocked turn from making a stray auxiliary compaction call.

The block isn't in a tagged release yet; the releases page shows v0.22.5, from 23 Oct, as the latest.

Measure Your Own Model Mix With Observer Hooks

GoldieBench tells you how models do on its tasks; your own logs tell you how they do on yours.

post_api_request fires after each successful provider call with the model, provider, duration, finish reason and usage.

api_request_error fires on each failed attempt with the status code and whether it's retryable.

Log both for a week and you'll see which model lane actually carries your traffic.

The field lists are in the observer hooks reference.

My Pairing, In One Table

Picking plugins for the same setup? Compare them in the best Hermes Agent plugins.

For the Agent OS with these lanes ready, join the AI Profit Boardroom, and for a plan built on your workload, book a free strategy session.

Also On Our Network

• Agent Operatorsthe six-step operator setup for shell hooks, fail-closed gates and turn blocks
• AI Income Deskhow a shop, a service firm and an agency use hooks to protect client data
• AI Tool Verdicta nine-criterion scorecard on control, safe defaults, consent and privacy
• agentos.guidefive copy-paste prompts that set up hooks inside your Agent OS
• aiprofitboardroom.comthe complete Hermes Agent Hooks guide on the AI Profit Boardroom blog

Pick your lane from the table, add the matching hook today, and if you want my team to check the pairing, book a free strategy session before you roll out Hermes agent hooks.