On Tuesday, September 1st, 2026 at 7:02 AM, kiako via agora-discussion
<[email protected]> wrote:
> Hi all,
>
> Given the 3(?) recent cases in Agora of LLMs being employed by players
> (automated and otherwise), the Discord's been discussing if we need a clear
> stance on LLMs and "agents" in Agora. To keep the conversation archived &
> accessible, I'm moving it to the list.
>
> I'd appreciate if those with vested interest in Agora could chip in their
> opinions.
>
> Some convo starters:
>
> * What do you think of LLMs in/for Agora?
> * What do you think of automated or "agentic" LLMs participating in Agora
> (those that would send messages without human prompting or vetting)?
> * Would allowing/prohibiting LLMs (automated/"agentic" or otherwise) affect
> your participation in Agora (positively or negatively)?
>
> I do ask that if comments are being generated or sent by an
> automated/"agentic" LLM workflow for this conversation, it should explicitly
> be stated as such.
>
> I'll be sending my thoughts in a separate message.
Here's my thoughts in a separate message:
I oppose LLMs in Agora. I've got some personal opinions about their
capabilities (or lack thereof), but I'd rather like to focus on the specific
issues of automated text output in Agora.
Using an LLM is an inherently antisocial behavior, which does not mesh well
with an intrinsically social game. The only value of an LLM, as currently
marketed and as primarily used, is
* to avoid reading text produced by humans,
* to avoid writing text needing to be read by humans,
* to avoid learning/practicing skills, and, in general,
* to avoid communicating with humans.
The purpose of an "agentic" LLM in Agora is to "play" the game without having
to engage personally with the game, at the cost of the other participants'
investments. I can marginally tolerate one "agentic" LLM - it is sometimes
interesting to see where it fails - but I lose interest once two LLMs generate
messages towards each other. Agora runs by the voluntary efforts of our
(absolutely wonderful <3) officers, so actions performed between LLMs are
actions initiated by no one, for the benefit of no one, that must be tracked by
someone.
The alternative is having LLMs generate reports, which, given their propensity
for errors, just kicks the can down the line to those reviewing and CoE'ing.
As a recent observation, we have a general (and quite important) cultural
practice of pointing out newer players' errors before those errors can bite
them. When an LLM generates text with errors, it's created busywork for humans
with no guarantee that future output will ever improve. For me, this is
incentive to not bother noting errors I've caught, creating a more hostile
culture for newer players to have to suffer through. I'd rather Agora appear
(and continue to be) friendly and welcoming towards all active participants.
As for non-agentic generative LLMs, I don't think they should be used. If you
find yourself reaching for one, I strongly encourage you to instead try the
following:
* If you're unsure about a rule, ask! We've got some incredibly
knowledgeable people with years (decades?) of experience in
interpreting Agoran rules, and an entire legal system for solving
the edge-cases.
* If you're looking for ideas or feel Agora needs something extra,
bring it up in DIS! Share your thoughts! Someone else almost
certainly feels the same way, so be the one to get the ball rolling!
* If you want proposal review or aren't sure how to translate an idea
to Agoran language, post it to DIS or the #drafting channel on
Discord/IRC/Matrix! The sooner & more eyes you get on a proposal,
the better chance it actually gets adopted at voting.
* If you're wanting to get in on scams, reach out to players on the
Discord, IRC, Matrix, or direct/cold e-mail. Figure out who's good
at scams and make sure they know you're available to collaborate.
* I am perpetually available to help come up with/implement automated
tooling, if you're just needing a way to process something in less
time.
We don't bite :)
kiako, Archivist, Webmastor