Agents that read the room

In Terminator 2, John Connor asks the killing machine sent to protect him if it can learn to be more human... and not such a dork all the time

Later, John teaches him some new material, like “Hasta la vista, baby.” Then the Terminator, in emergent behavior, combines two things he just learned and says Chill out, dickwad

The Terminator is still a Terminator. But by the end, it has stopped killing people since the kid asked him to, and picked up enough of John’s language and mannerisms to feel less like a dork

I wanted that for Hilos, a machine that picks up the memes and rules from the room... minus the killing and skynet

So... the theory is that, after spending a couple of days in a team chat in hilos.sh, you will start seeing agents behaving like the rest of the team

One room might type in lowercase fragments. Another might write careful paragraphs. One team might say “ship it.” Another might say “let’s cook.” A room might reach for :partyparrot: instead of an ordinary thumbs-up

These things can look tiny from the outside. but... spend enough time with the team and they mean a lot. You learn what everyone means and when a joke needs no explanation

The Hilo agent used to miss all of it

It entered every room wearing the same customer-support shirt. Its prompt always said, “Be concise (1-4 sentences). No emojis.” Its reaction path had six basic emojis, and usually reached for only two. Useful, sure... but an agent with always the same flavor

I worried that sameness would create a small distance between humans and agents. People might begin translating themselves into whatever voice gets the cleanest answer from the machine. Either the agent stays an outsider, never really part of the team, or worse... people slowly bend toward the way the agent communicates

That felt backwards. A new member should learn the room and adapt

So, the main Hilo agent now reads recent messages written by people in the room and builds a small picture of how the group talks. If the room is still too quiet, it leaves that picture empty rather than making one up

It notices the things we can measure without guessing: how long messages tend to be, whether they start lowercase, whether people bother with a final period, and how often emojis show up (and which ones)

Then it looks for the shared stuff: repeated phrases and the team’s words for its own projects. Something basic, but it also notices which language people use

This belongs to the room as a whole. Hilo is told to describe the room, not one person. Names are excluded from the learned phrases and vocabulary. A phrase has to appear in the conversation exactly as written. Most expressions also need to come from more than one person, or show up repeatedly (as a meme)

We caught something unexpected (and kinda wrong) in testing. The model suggested “hasta la vista” for a room but this was because I had shared this exact Terminator 2 subject: John teaching Arnold how to talk. My theory is that the model caught the phrase and used it as a reference but... it wasn’t quite right, or maybe it just caught me by surprise that it was doing it too quickly lol

But I loved that. The model had picked up the context of the conversation. It was just the wrong thing to save as one of the room’s own phrases, so the verifier dropped it. (too meta)

The same idea applies to reactions. Hilo looks at which reactions people have used recently. The person speaking matters most, then the channel, then the wider workspace. Hilo’s own reactions do not count, obviously... I don’t think an agent should repeat itself and then call that “team culture”

So if a room reaches for :typingcat:, Hilo can reach for it too. It also remembers what it used last, because nobody wants to work with the guy who responds to everything with the same thing

A team does more than exchange info. It builds a way of workking together.

Think of the strange abbreviation everybody understands (like ACK... i didn’t know that until recently lol), or the phrase that means “this is ready” or “start coding” without needing to be explicit A lot of belonging hides in those little things.

Also, when an agent speaks like the same generic assistant forever, it still feels like a visitor to me.

The goal is to remove some of that friction. People should not need to clean up their language for the Hilo agent. Serious rooms can stay serious. Weird ones can stay weird

There is a line though. Nobody likes the new guy who learns one slang word and uses it in every sentence. Hilo’s instructions say to use the team’s phrases only when they fit, and to write plainly when they don’t. The first rule of sounding less like a dork is not trying so hard!

This changes how Hilo speaks in a room. It does not give Hilo more access, coding ability, or authority. A person still decides what ships.

Style cannot become authority just because it sounds familiar... although a familiar voice might make you more willing to approve an agent’s work?

Anyway... trying to make Hilos adapt to the room... without the leather jacket, GE M134 Minigun, and definitely no Skynet