Two Hermes Agent commands, three different answers

Setting up /bg and /btw, in about five minutes#

Start with the setup, because the two commands are worth having working before you care why they confuse people. Both ship with Hermes Agent: nothing to install beyond the agent itself, and no plugin to enable. What follows is drawn from the project's own documentation, checked this morning.

That is the whole setup. The Discord gateway documentation lists both commands among the built-ins, and notably /bg is available on the Discord gateway too, not just the terminal, which matters if you want work moving while you are nowhere near a keyboard.

Two configuration notes worth knowing before you scale it. Group sessions are separated per user by default, so in a shared channel each person gets their own session rather than inheriting somebody else's history. And scheduled work is handled by a separate cron command, with subcommands to list, create, edit, pause, resume and remove jobs, plus a per-job reasoning setting.

With that working, the rest of this piece is about the part that trips teams up: what each command actually does with your context, and why three different sources currently give three different answers.

The week the commands swapped coats#

On Saturday the maintainer of Hermes Agent posted a short note that quietly rearranged two of its most-used commands. Nous Research, writing as Teknium, said the feedback had landed and that background work was being re-cut: the fresh-session behaviour moved to /bg, and /btw was changed to fork the session you are sitting in. The post has been seen more than thirty thousand times.

Those counts come from the release notes themselves, not from a tracker, and each one states the window it measures. Version 0.20.1, tagged on 13 August, rolled up about 656 merged pull requests since v0.20.0. Version 0.20.2 followed on 16 August with roughly 397. Version 0.20.3 landed the same day at about 125. Version 0.20.4, on 18 August, was a small one at roughly 74. Version 0.20.5 arrived on 19 August with about 323. Version 0.20.6, tagged on 27 August, carried roughly 525 merged pull requests and about 1,313 commits across some 1,557 files.

One figure is deliberately absent from that chart. Version 0.20.0, on 3 August, reports roughly 3,650 commits and about 1,400 merged pull requests, but it measures back to v0.19.0 rather than to the release before it. Plotting it beside the patch windows would draw a spike that is an artefact of the baseline, not of the work, so it sits here in the text instead.

Read that as a scent trail rather than a scoreboard, and follow the tracks rather than the summary. A project moving at this pace will outrun any second-hand summary of it, and the commands that changed on Saturday are exactly the sort of small surface that shifts without a headline. Anyone building a repeatable ai agent workflow on top of it is building on ground that moves.

One reply under that post did the job of a whole support queue. A user going by Creed Hardcastle wrote that half the background bugs are people who wanted a clean room and got a clone. That single sentence is the whole confusion in miniature, and it is worth keeping in mind for the rest of this piece.

What Hermes Agent does with /bg#

The official slash command reference describes /bg plainly: it runs a prompt in a separate background session, the agent works through it independently, and your current session stays free for other work. Results arrive as a panel when the task finishes.

The CLI guide goes further, and the detail there is the part that matters. Each /bg prompt spawns a completely separate agent session in a daemon thread, holding an isolated conversation. The background agent receives your prompt and nothing else. It does not inherit the history of the chat you fired it from. What it does inherit is configuration: your model, provider, toolsets, reasoning settings and fallback model all carry across.

Several can run at once. Each gets a numbered identifier, in the shape bg_143022_a1b2c3, and the documentation sets no explicit ceiling on how many you may have in flight. Your foreground session, the guide says, stays fully interactive: you can chat, run commands, or start further background tasks while the first ones are still working.

Two smaller details deserve a mention because they change how you would build around it. A finished task can ring the terminal bell if you switch that on, and background sessions never appear in your main conversation history. They are standalone, with their own task identifier. That is a clean separation, and a quiet one: the work happens somewhere you are not looking.

So the correct mental model for /bg is a courier, not a colleague. It leaves the den with your brief and does not come back to ask. You hand it a sealed brief. It cannot ask what you meant, because it never saw the conversation where you worked out what you meant. Brief it as you would brief a freelancer who has not read the thread, which is the same discipline good content operations work has always needed.

What /btw does while the turn keeps running#

The same reference page describes /btw as a way to ask a quick side question about the current conversation without interrupting it. Per that page, a one-shot auxiliary model call answers from a read-only snapshot of the transcript, the live session's history and prompt cache are left untouched, and the current turn carries on working.

Set the two side by side and the shape is almost a mirror. A background agent started with /bg sees none of your session. A side question asked with /btw sees only your session. One is sealed off from the context; the other is made entirely of it. They are not two names for one behaviour, and treating them as interchangeable is how you end up with the clean room that turned out to be a clone.

Here is where honesty matters more than tidiness. The reference says read-only snapshot. The maintainer, on 29 August, said /btw will fork your session off in the background. A fork and a read-only snapshot are not the same promise. A snapshot reads; a fork continues. This piece is not going to pretend those two descriptions have been reconciled, because as of this morning they have not been.

Three sources, three answers#

The third description is the one most people will actually meet, and it is the one that is plainly wrong. A widely syndicated command reference, the Toolsbase cheat sheet, lists the entry as /background and then gives its aliases as /bg and /btw. On that page the two are the same command wearing two names.

Why does a wrong cheat sheet matter enough to write about? Because it is the artefact most teams standardise on. Nobody circulates a link to a documentation subpage in a team channel. They circulate the one-page table, it gets pasted into an onboarding doc, and a year later it is still there, describing a product that has moved several times since.

That is the durable lesson, and it travels well beyond one project. Documentation for a fast-moving tool is a snapshot, not a contract. The same caution applies to the rules governing agent access to your site and to every vendor policy page a media team quotes back at a client.

How to brief an agent you cannot watch#

None of this is a reason to avoid background work. It is a reason to write the brief properly. The cost of an interruption is well studied, and the finding is sharper than the folklore suggests.

In The Cost of Interrupted Work, CHI 2008, Gloria Mark, Daniela Gudith and Ulrich Klocke reported something counterintuitive from forty-eight participants: people completed interrupted tasks in less time, with no difference in quality. The cost did not show up on the clock. It showed up in the people. Their published finding is that participants compensated by working faster, and paid for it in stress, frustration, time pressure and effort.

That reframes what ai agent workflow automation is actually buying you. It is not primarily speed. It is the removal of the strain that comes from holding two threads at once. A task handed cleanly to a background session is one you have genuinely put down, which is a different thing from one you are half-watching.

Two operational habits follow. First, pin the version. If a process depends on what a command does, record which release you tested against, because this project alone shipped five tagged releases in August. Second, brief for isolation by default: write the background prompt as though the reader has seen nothing, since under the documented behaviour that is precisely true.

Hermes Agent is not unusual in any of this. It is simply moving fast enough and publishing openly enough that you can watch the seams. Most tools a marketing team leans on are doing the same thing behind a changelog nobody reads, which is the quiet risk in every search and answer-engine programme built on somebody else's platform.

The vulpine move here is a patient prowl rather than a pounce. Check the primary source, note the date beside it, and keep a short trail of what you verified and when. That habit costs a few minutes and saves the strange afternoon where a process breaks because a command quietly changed coats on a Saturday.

Frequently asked questions#