Skip to content
vavio

Can Hermes keep working on a goal without me prompting it?

Yes. Send /goal followed by what you want done, and Hermes keeps taking turns in the same conversation until a judge model decides the goal is met, decides it cannot be met without you, or the turn budget runs out. The default budget is 20 turns, after which Hermes pauses and asks you to /goal resume or /goal clear. Each turn costs what an ordinary turn costs, plus one judge call, so a long goal costs roughly what the same number of “keep going” messages would, plus the judge.

This follows Hermes Agent v0.23.1, commit 3b42157, checked on 17 November 2026. Upstream documents the same /goal command for the terminal, the desktop and dashboard chat, and messaging channels such as Telegram. We reviewed official documentation and source. We did not run a goal on a Vavio computer or measure completion rates or cost.

How do I start a goal?

Write the goal as an outcome Hermes can show it has reached:

/goal Collect this week's invoices from the shared folder into one spreadsheet, total them, and reply with the total and the file path

Hermes replies ⊙ Goal set (20-turn budget) and starts the first turn straight away, so you don’t send a separate message. After each reply you see either ↻ Continuing toward goal (3/20): <reason>, ✓ Goal achieved: <reason> or a pause notice. The reason is the judge’s one-sentence explanation, which tells you what it saw.

Command What it does
/goal <text> Sets or replaces the goal and starts working
/goal or /goal status Shows the goal, its status and turns used
/goal pause Stops the loop but keeps the goal
/goal resume Restarts the loop and resets the turn count to zero
/goal clear Drops the goal
/subgoal <text> Adds another condition that must also be met

There is one goal per conversation. While a turn is running, control commands such as status, pause, resume and clear are safe to send. Setting a new goal is refused until you /stop the current turn.

What does the judge decide?

After every turn, Hermes sends the judge your goal, the last 4,000 characters of Hermes’s reply and a list of any background processes Hermes started. The judge returns one of four verdicts:

  • done: the reply explicitly says the goal is complete or shows the finished deliverable. The loop stops.
  • blocked: the reply says the goal is impossible or needs your input, such as a sign-in or a decision. The goal pauses with the reason, so it doesn’t spend turns on a task it can’t finish.
  • wait: Hermes is waiting on something it started, such as a build or a download. The loop parks without using turns or judge calls until that process ends or a delay passes (at most 30 minutes for a process). In the terminal the loop then resumes by itself. In the pinned source, the messaging gateway only checks again on the next turn, so in a chat it may stay parked until a message arrives: send /goal unwait or any message to wake it.
  • continue: there is a next step. Hermes adds “[Continuing toward your standing goal]” to the conversation and takes another turn.

The judge reads what Hermes said, not the result itself. A reply claiming a file exists can be judged done even if the file is wrong. To make “done” mean more, put a check on its own line under the goal (verify: the spreadsheet has one row per invoice). The judge then requires concrete evidence, though it still runs nothing itself. For a mechanical check, add a quality gate: /goal gate add <command> runs a shell command before each verdict, and the goal can’t complete until it succeeds. In a messaging channel, only a configured gateway admin can add gates.

If the judge itself fails, Hermes treats that as continue. It pauses the goal after three unreadable verdicts in a row or five failed judge calls in a row, such as a wrong API key, rather than spending the whole budget.

What stops a goal?

Stop What you see To carry on
Judge says done ✓ Goal achieved Nothing to do; send a new /goal for more
Judge says blocked 🚫 Goal judged unachievable — paused Provide what it needs, then /goal resume, or reword with /goal <text>
Turn budget used ⏸ Goal paused — 20/20 turns used /goal resume gives another full budget
Judge keeps failing ⏸ Goal paused, naming the judge Fix the judge model setting, then /goal resume
A turn fails or returns nothing, such as a model error The error, and no continuation In a chat the goal stays active but idle (the terminal pauses it after an interrupt): fix the cause, then /goal resume
A quality gate keeps failing past its retries (3 by default) ⏸ Goal paused, naming the gate Fix the problem or remove the gate, then /goal resume
You pause or clear it Confirmation /goal resume after a pause

In the pinned source the budget counts every turn while the goal is active: the first turn, each automatic continuation and any message you send yourself. So with the default of 20, Hermes takes at most 20 turns before pausing, not 20 extra turns. Upstream’s guide calls them continuation turns; we followed the source. Change the limit with goals.max_turns in the Hermes profile’s config.yaml.

Your own messages are not skipped. In a messaging channel, the next continuation joins the queue behind anything you have already sent, and the judge runs again after your turn, so a message that finishes the job also ends the goal.

How much does a goal cost?

A goal turn costs the same as a normal turn of the same size: Hermes calls your main model with the conversation and tools, often several times within one turn, and the continuation is an ordinary message, so prompt caching works as usual. On top of that, each turn normally has one judge call (a failing quality gate skips it). The judge sees a fixed instruction, the goal and up to 4,000 characters of the last reply, not the whole conversation. By default it uses your main model, and the pinned source allows it up to 4,096 output tokens so that reasoning models have room to answer; upstream’s guide estimates about 200. On an expensive reasoning model, the judge is not a rounding error.

Two settings bound the bill:

  • goals.max_turns caps model turns before you’re asked again. Lower it while you’re learning how a model behaves on your tasks.
  • auxiliary.goal_judge can route the judge to a cheaper, fast model that reliably returns the short JSON verdict it needs.

Remember that each turn re-sends a growing conversation, so later turns cost more than early ones. The long-conversation billing guide explains why. Check prices with your model provider.

Is /goal the right tool, or do I want a schedule?

You want Hermes to Use How it runs
Keep working on one task until it’s finished /goal Turn after turn in this conversation, until done, blocked, paused or out of budget
Repeat a check every few minutes while you’re in a conversation /loop 5m <prompt> On a timer inside this conversation
Do something at set times, such as a morning summary A scheduled (cron) job In a fresh session each time, results delivered to a chat or file
Do one separate task without tying up this chat /bg <prompt> Once, in its own background session, with the result sent back to the chat

A goal does not schedule anything. It stops when the judge says done or blocked, and it does not wait until tomorrow. For recurring work, see whether a job uses another job’s result and how to stop a running job without deleting its schedule.

What happens to a goal when Hermes restarts?

The goal itself is saved with the conversation, including its status, turn count, conditions and gates. The pending continuation is not. In the messaging gateway, Hermes queues the next continuation in memory after delivering each reply, not in the saved message queue. So after a restart, send /goal status. If the goal is still active but Hermes has stopped working on it, /goal resume starts the next turn and resets the budget. The restart guide covers a turn that was running when Hermes stopped. We didn’t test a restart during a goal.

What hosting changes

A goal is most useful when you start it and then leave: send the goal from your phone, and Hermes keeps working while your laptop is closed. That only works if the computer Hermes runs on stays on and connected. Vavio supplies a managed Hermes computer for $14.99 per month, with model usage billed separately. Hosting keeps Hermes running. It doesn’t change how the judge decides, what the budget allows or what each turn costs. Review the current hosting offer if that is the service you need.

Sources

Written by Vavio with AI assistance and checked against pinned official Hermes v0.23.1 documentation and source. This is a source review, not a live goal, completion or cost trial.

More from the Vavio blog: Hermes Agent guides