Type /btw followed by a question and Hermes agent /btw answers it on the side, without interrupting the task that's running — a one-shot call against a read-only snapshot of the transcript, so the live session's history and prompt cache stay exactly as they were while the current turn carries on. I run Hermes agents across my SEO business every single day, and this small command has quietly fixed one of my worst habits: derailing a perfectly good long-running job because I wanted to ask something trivial.
📺 Watch: Hermes Agent OS Just Changed AI Agents Forever!
🔥 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
Before /btw existed, every option was bad. Interrupt the task and pay for it. Open a second session and lose the context. Or sit on my hands until the job finished. In this guide I'll cover what the command actually does, why the untouched-cache detail matters far more than it sounds, and how it differs from the two things people keep confusing it with.
What the Hermes agent /btw command actually does
The syntax is /btw <question> — that's the whole thing. The official slash-commands reference describes /btw as asking a quick side question about the current conversation without interrupting it. Under the bonnet, a one-shot auxiliary LLM call answers from a read-only snapshot of the transcript. The live session's history and prompt cache are untouched, and the current turn keeps running the entire time.
Here's my plain-English version. Picture the agent's conversation as a document someone is actively writing. /btw photocopies that document as it stands, hands the copy to a second reader in another room, and that reader answers your question from the copy. The writer never stops typing, never sees the question, and the original document doesn't gain a single word. You get your answer, and the task never knew you asked.
/btw works in the CLI and through the messaging gateway too, so the behaviour is identical whether I'm at my desk or checking on an agent from my phone between calls.
Why the untouched cache matters more than it sounds
"The prompt cache is untouched" reads like a footnote. It's actually the headline. As I broke down in my Hermes memory budget guide, agents re-read their entire context on every single step, and a stable context caches cheaply. The moment something new enters the conversation, the cache behind it stops being valid, and every step that follows pays to rebuild what used to be nearly free.
That's why a casual mid-task question is never just one question's worth of cost — it's a wedge driven into the transcript that every remaining step has to re-process. Hermes agent /btw is that caching principle turned into a command: the question and its answer live entirely outside the session, the cache stays warm, and a fifty-step job finishes at the same cheap rate it started at. Curiosity stops carrying a tax.
Want the full command playbook I actually run my agents on — /btw included, with the workflows the docs don't give you? Join the AI Profit Boardroom and steal my Hermes agent workflows. Prefer to talk it through first? Book a free SEO strategy session and I'll map it onto your setup.
📺 Watch: Hermes3D Is INSANE: AI Agents Now Have a 3D Office
/btw vs /bg: the mix-up half the community made
The docs draw the line in one sentence: /btw is for a quick side question about the current conversation, and for independent work with fresh context you use /bg instead. One command reads over the transcript's shoulder; the other spins up separate work that doesn't need the transcript at all. I've covered the Hermes /bg command in its own guide, and the two are opposites wearing similar clothing.
The community is only just catching on. The unofficial Hermes Release Watch cheat sheet from 5 September 2026 listed /btw, and one user admitted they'd been using /bg for side questions the whole time — only now learning why /btw is the accurate choice. It's an easy mistake with a real cost: /bg starts with fresh context, so a background agent doesn't know your conversation and can only guess at questions about it. /btw exists precisely because it can see the conversation. Wrong tool, plausible-looking output, quietly worse answers.
To be fair, releases are landing faster than the knowledge spreads — my Hermes v0.21 breakdown shows how much ships in a single version — so I don't blame anyone for missing a command. But this pair is worth getting right, because the failure mode is silent.
/btw vs just messaging the agent mid-task
The other habit /btw replaces is typing a normal message while the agent is working. A plain message lands inside the live session, which means Hermes has to deal with it as part of the job — I've written up how Hermes handles messages while it's busy separately, and the short version is that your "quick question" becomes the agent's problem to manage.
/btw never enters the session at all, so there's nothing for the agent to manage. My rule of thumb is simple: if the message should change what the agent does, it belongs in the session. If it's purely for my benefit — context, clarity, curiosity — it's a /btw. Getting that split right is half the skill of supervising agents without slowing them down.
📺 Watch: Hermes AI Agents Just Went Portable
Practical ways I use /btw every day
Once it stops feeling like a novelty, /btw becomes the command you reach for constantly. My most-used patterns:
- Check what the agent meant. Forty minutes into a job it references "the earlier approach" — /btw which approach are you referring to, and I get clarity without touching the task.
- Ask where it's up to. /btw how far through the article batch are we — a progress check that costs the job nothing.
- Clarify a term. Long transcripts fill up with project shorthand — /btw what does the staging bucket refer to here.
- Sanity-check before intervening. I ask the side question first, and half the time the answer proves the agent is on track — so the interruption I was about to make never happens.
Inside my Agent OS setup, long multi-step jobs are the default rather than the exception, and /btw is a big part of what makes them tolerable to supervise. When a side answer turns out to be worth keeping permanently, I feed it back properly once the task finishes — that's a job for the Hermes /learn command. /btw for the moment, /learn for the long run.
The auxiliary-model angle
Notice the exact wording in the docs: a one-shot auxiliary LLM call. Your side question isn't answered by the main agent at all — it's handled by an auxiliary call, the same pattern I dug into in my roundup of the best auxiliary models for Hermes. That design is what makes /btw feel free: the heavyweight model stays locked on the task while the one-shot call does the reading and answering on the side.
It's also the sort of claim I test rather than take on faith — Goldie Bench is my standing rig for putting every new Hermes command through its paces before it earns a place in my workflows. /btw passed quickly: side questions answered, main task undisturbed, exactly what the reference promises.
Which command do you actually need?
| What you need | What to reach for |
|---|---|
| A quick side question about the conversation that's running | /btw <question> |
| Independent work with fresh context | /bg |
| To change what the agent is actually doing | A normal message in the session |
| To keep a lesson for future sessions | /learn |
FAQs about Hermes agent /btw
Does /btw interrupt the task Hermes is running?
No — that's the entire point. The current turn keeps running while a one-shot call answers your question from a read-only snapshot of the transcript. The session never pauses and never sees the exchange.
Does the /btw answer become part of the conversation?
No. Because the answer comes from a snapshot rather than from inside the session, the live history and prompt cache stay untouched. If the answer reveals something the agent genuinely needs to know, send that as a proper message afterwards — /btw itself changes nothing.
Does /btw work outside the CLI?
Yes. It works in the CLI and in the messaging gateway, which is where I use it most — quick side questions from my phone while an agent grinds through a batch job back home.
When should I use /bg instead of /btw?
Follow the docs' own contrast: /btw for a quick question about the current conversation, /bg for independent work that wants fresh context. If the question only makes sense with the transcript in view, it's /btw territory.
Which model answers a /btw question?
An auxiliary one-shot call handles it rather than the main agent, which is why the running task doesn't slow down. Choosing that auxiliary layer well is its own topic — see my auxiliary models guide linked above.
My verdict on /btw
Hermes agent /btw is the smallest command in the reference and one of the most used in my day. It answers the question you were going to ask anyway, minus the cache invalidation, the derailed turn, and the pointless second session. Learn the split — /btw to understand, a message to steer, /bg to delegate — and supervising long agent runs stops feeling like defusing a bomb.
If you're building towards agents that run for hours while you ask questions from the sidelines, that's exactly what I teach — join the AI Profit Boardroom and get my full Hermes agent stack, or book a free SEO strategy session and we'll plan your rollout one-on-one.
