Delegating to an open source AI agent while you raid

Setting up the handover, in six steps#

Setup first, argument later. The goal for an open source ai agent is narrow and testable: work leaves your hands, runs while you are elsewhere, and the result lands somewhere you will actually see it. Everything below is from the public repository of Hermes Agent, an open source ai agent from Nous Research, read this morning.

Two configuration details are worth knowing before more than one person uses 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 there is a ceiling of one hundred application commands per bot on Discord, which only bites once you are installing many skills.

One honest caveat on the scheduling step. The scheduled tasks documentation shows scheduled jobs being managed from chat, and says plainly that no command line is required. The slash command reference lists the same command as terminal-only. Two official pages, two answers, so test it on your own install rather than trusting either.

Why a raid night tests an open source AI agent honestly#

Most delegation is tested badly, because most of us never fully leave. We hand something over and then hover, which measures nothing. A raid night is different. It is scheduled in advance, it runs for a fixed stretch, other people are depending on your attention, and you cannot quietly alt-tab back to check on a job without letting someone down.

That makes it a genuinely useful proving ground for any ai personal assistant. If the handover works across a raid, it works across a long meeting, a flight, a school run or a holiday. If it only works when you are half-watching, it was never a handover at all.

Gamers worked this out before marketers did. Anyone who raids already runs a second screen of tooling, already lives in a voice channel, and already accepts that some things happen on a schedule rather than on demand. Pointing an ai agent automation setup at that same channel is a small step, not a leap, and the habits transfer straight back into working hours.

The quiet prize is not the raid. It is discovering which of your tasks were genuinely separable all along, and which only felt urgent because you were sitting next to them.

There is a second reason the gaming frame earns its place. Raiders are unusually honest about tooling, because a broken addon shows up immediately and publicly. Marketing teams are not: a workflow can quietly fail for weeks before anybody picks up the scent. Setting an open source ai agent loose against a hard deadline you cannot fudge is a fast way to find the weak link in your own trail.

Start small and keep the quarry modest. One task, one channel, one evening. The temptation with any open source ai agent is to hand it the whole den at once, which guarantees you cannot tell which part failed. A single job that reports back cleanly is worth more than six that half-run.

What interruption actually costs#

There is a number everyone quotes about interrupted work, and it is worth putting down carefully, because the paper it is usually attached to does not contain it.

In The Cost of Interrupted Work, CHI 2008, Gloria Mark, Daniela Gudith and Ulrich Klocke ran an experiment with forty-eight participants. Their reported finding is not that interrupted people were slower. It is the opposite: participants completed interrupted tasks in less time, with no difference in quality. What they paid instead was stress, frustration, time pressure and effort, because they compensated by working faster.

Read that again, because it changes the case for handing work away. The reason to point an open source ai agent at background work is not that you are slow when interrupted. It is that interleaving is expensive in a currency the clock never shows, and ai productivity tools that only shave seconds are solving the wrong problem.

Cite it for what it measured and the argument gets stronger, not weaker. A background session removes the interleaving entirely: you are not switching between two threads, you have genuinely put one down. That is the whole benefit, and it survives scrutiny in a way the folk statistic does not.

What is safe to hand over, and what is not#

Not everything is separable. The test is simple: could a capable stranger who has not read your afternoon do this from the brief alone? A background session receives only the prompt you send it, so that stranger is exactly who you are briefing.

The habit that makes this work is writing the brief as though you will not be available to clarify it, which is precisely true. It is the same discipline good content operations already demands, and the same one that makes a handover to a human colleague survive a week off.

A useful way to sort the pile is to ask what a colleague would need in writing. If you can put the whole job in one message with no follow-up, an open source ai agent can take it. If your answer starts with 'well, it depends', that dependency is context you have not yet written down, and no amount of model quality will recover it from thin air.

Version discipline belongs here too. The release tagged on 19 August and the one before it tagged on 13 August both moved substantial ground, so the behaviour you tested a fortnight ago is not necessarily the behaviour you have now. Pin the release you verified against and note it beside the job.

Beware second-hand instructions in particular. A widely circulated third-party cheat sheet currently describes two of these commands as aliases of one another, which the project's own reference contradicts. Any open source ai agent worth running will be moving faster than the guides written about it.

Start with the jobs you would happily automate repetitive tasks away from a Monday morning: the recurring pull, the standing check, the tidy-up. Leave anything where the cost of being wrong lands on a client. That boundary is the same one folkfox draws around every search and answer-engine programme: automation handles the gathering, people own the call.

Where the handover breaks#

Defaults decide whether any of this reaches you, and they are published. The scheduled tasks documentation states how long the agent waits before giving up on each kind of work, and the spread is wide enough to matter.

The command reference also documents the surfaces this runs on: a gateway that can be installed as a background service, a backend server for desktop and remote clients, and a one-shot send command intended for ops scripts, scheduled jobs and monitoring. That last one matters because it delivers a message without spinning up an agent loop at all, which is often exactly what you want at three in the morning.

Three more defaults are worth setting deliberately. The scheduler wakes every sixty seconds, so nothing is instant. Three consecutive failures trigger a review nudge rather than silence. And a ten-minute grace window governs how long the system waits before catching up on a job it missed.

The honest limits are worth stating plainly. A background session cannot ask you what you meant. Its results do not join your main conversation history, so they need somewhere to land. The commands themselves moved as recently as 29 August, and the release tagged on 27 August carried roughly 525 merged pull requests, so anything you verify today is verified against today's build.

The failure mode to watch for is silence that looks like success. A task that outlived its timeout leaves no trail in your channel, and an open source ai agent that has quietly given up looks identical to one still working. That is why the rehearsal step is not optional: you are not testing whether it can do the job, you are testing whether you find out when it cannot.

None of that is a reason to keep everything in your own paws. It is a reason to test the trail once, in a block of time where you genuinely cannot intervene, and to write down what you pinned. Come back from the raid, read the panel, and you will know whether the handover holds. The sister piece on those two commands of this newsroom's own tooling gets the same treatment.

Frequently asked questions#