Type /review at the end of a working session and Hermes agent /review spawns an independent, full-privilege reviewer subagent that digs into whatever you have just been discussing — a pull request, a block of code, documentation, any artifact referenced in your last ten chat messages — and reports back into your session. It investigates in the background, opening the PR, reading the diff and running code where it needs to, and the finished review re-enters your session as a completion your primary agent can act on straight away. One command, one fresh pair of eyes, and no copy-pasting context into a second window.
📺 Watch: Hermes Bot Mode DESTROYS Grok Bot? (FREE!)
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
I have been running it on every meaningful piece of work since it landed, so in this guide I will cover exactly what the docs say the command does, how to steer it with instructions, how to pin a dedicated reviewer model with auxiliary.review, and where it slots into a plan, build, review loop.
What the Hermes agent /review command does
The documentation files /review under Subagent Delegation, and that placement tells you most of the story: Hermes hands review duty to a separate subagent instead of marking its own homework. Three documented details matter.
- It is independent and full-privilege. The reviewer is its own subagent with real access, so it genuinely investigates — it opens the PR, reads the diff and runs code rather than skimming a summary of the conversation.
- It scopes to the work you have just been discussing. Anything referenced in the last ten chat messages is fair game: a PR, code, docs, any artifact the conversation has touched.
- The review comes back to you. The subagent works in the background and its full review re-enters the session as a background-subagent completion, which your primary agent can then act on — fixing findings, pushing changes, or challenging the verdict.
Per the official slash-commands reference, /review works both in the CLI and through the messaging gateway, so the same review loop runs whether I am at my desk or firing it from my phone. The command was also flagged on the unofficial Hermes Release Watch cheat sheet on 5 September 2026, which is where a lot of people first clocked it.
Syntax and steering: /review with optional instructions
The documented syntax is deliberately minimal: /review on its own, or /review followed by optional instructions. Bare /review hands the reviewer your recent context and lets it decide what matters. Adding instructions is where the command earns its keep, because you are effectively writing a one-line review brief for an independent specialist.
Three ways I steer /review in practice:
- Focus on one axis. Error handling, security, rollback safety — when I already trust the rest of the change, I tell the reviewer where the risk lives.
- Review against intent. I restate what the change was supposed to achieve, so the subagent judges outcomes rather than style points.
- Ask for ranked findings. Severity-ordered output means my primary agent fixes the dangerous items first instead of tidying trivia.
The tighter the brief, the sharper the findings that land back in your session — vague requests get vague reviews, same as with a human reviewer.
Want the exact review briefs I feed /review, plus my full Hermes command playbook? Grab my Hermes review workflows inside the AI Profit Boardroom — and if you would rather map it to your own stack one on one, book a free SEO strategy session and I will walk you through it.
📺 Watch: Hermes3D Is INSANE: AI Agents Now Have a 3D Office
Pinning a dedicated reviewer model with auxiliary.review
By default the /review subagent runs on your main model. The docs let you change that: set auxiliary.review in config.yaml and Hermes pins a dedicated model for review jobs while your primary model keeps building. This is the same pattern I broke down in my guide to the best auxiliary models for Hermes — different brains for side jobs — except review is the one side job where I spend up rather than down, because a reviewer that misses a real bug costs far more than the tokens you saved.
Choosing that review brain is worth actual testing. When I ran candidate models through Goldie Bench, the gap between them on spot-the-defect style tasks was much wider than on generation tasks, so the default is not automatically the best pick. And if you change nothing, /review still works fine — it simply inherits whatever your main model is.
Why an independent /review beats self-review
Ask an agent to check its own work and it mostly defends its own work. An independent reviewer starts with none of the builder's assumptions, which is why independent review has been a standing quality gate in my Hermes agent best practices since long before there was a dedicated command for it.
What /review changes is the friction. The old version of this discipline meant opening a second session, pasting context across by hand, and stitching the verdict back into the original conversation. Now the entire loop is one command: fresh eyes, full privileges, findings delivered into the same session where the fixes will happen. Quality gates only survive when they are cheap enough that you never skip them, and /review makes this one close to free.
📺 Watch: Hermes AI Agents Just Went Portable
Where /review fits: plan, build, review
On its own, /review is a handy checker. Inside a structured loop it becomes the third leg of the workflow I run daily: scope the work with the /plan command, let the agent build, then run /review before anything ships. When the reviewer flags rough edges rather than outright defects, I follow up with /refine to polish exactly what it flagged, so nothing gets reworked on vibes.
A lot of my review habits actually started on the Claude Code side, and if you run both tools the loop transfers cleanly — I covered the pairing in my Hermes agent and Claude Code guide. Inside my Agent OS setup, /review is simply the mandatory gate between the build agent claiming it is done and the work actually shipping.
Steering the reviewer live: the v0.21 angle
Because /review runs as a background subagent, it benefits directly from the live subagent steering that arrived in Hermes v0.21. The reviewer is not a black box while it works: if I realise mid-investigation that the risky part is the database migration, I can steer the running subagent towards it instead of waiting for the full review and then running /review a second time with a new brief. Steering plus independent review is a genuinely different experience from fire-and-forget checking.
Hermes agent /review at a glance
| Aspect | Documented behaviour |
|---|---|
| Syntax | /review, with optional instructions to steer the reviewer |
| What it spawns | An independent, full-privilege reviewer subagent |
| Scope | The work just discussed — a PR, code, docs, any artifact referenced in the last ten chat messages |
| How it investigates | In the background: opens the PR, reads the diff, runs code |
| How results arrive | The full review re-enters the session as a background-subagent completion the primary agent can act on |
| Reviewer model | Defaults to your main model; pin one via auxiliary.review in config.yaml |
| Documented under | Subagent Delegation |
| Where it works | The CLI and the messaging gateway |
/review FAQ
Does /review need any setup before I use it?
No. Out of the box, /review spawns the reviewer subagent on your main model and scopes itself to the last ten messages automatically. The only optional configuration is pinning a dedicated review model through auxiliary.review in config.yaml.
What can the /review subagent actually do while reviewing?
It is full-privilege, and the docs are explicit about what that looks like in practice: it opens the PR, reads the diff and runs code as part of a background investigation. You get a reviewer that has executed the work, not one that guessed from a description.
Does /review only work on code?
No. The documented scope is the work you have just been discussing — a PR, code, docs, or any artifact referenced in the last ten chat messages. I use it on written deliverables almost as often as on diffs.
Can I run /review away from my desk?
Yes. The command works in the CLI and through the messaging gateway, so you can trigger a full review from chat and have the completion waiting in your session.
How is /review different from /refine?
The way I use them: /review diagnoses, /refine treats. Review spawns an independent subagent to investigate and report findings; refine is what I reach for afterwards to improve the work itself, usually pointed at exactly what the reviewer flagged.
My verdict
Hermes agent /review is the cheapest quality gate I have ever added to a workflow: one command buys you an independent, full-privilege reviewer that investigates in the background and hands its findings to the agent best placed to fix them. The two upgrades worth making early are writing proper review briefs instead of firing it bare, and testing which model deserves the auxiliary.review pin. Do both and self-review quietly disappears from your pipeline — which is exactly where it belongs.
If you want the full system — my review briefs, the reviewer-model pins I actually run, and my plan, build, review templates — join me inside the AI Profit Boardroom and steal my /review workflow. Prefer a personal walkthrough? Book a free SEO strategy session and we will wire it into your setup together.
