Sydney Runkle uses Jev to decide if her Slack agent replies
The gate runs before the agent speaks, and the case she names is threads where humans are talking to each other.
Sydney Runkle says her team's software factory Slack app uses Jev to decide if the agent should respond. The call sits in front of the agent rather than inside it. Jev is not writing the reply, it is deciding whether there is a reply at all.
people are often just talking to each other and the agent should only jump in when someone actually loops it in
The case she singles out is multiplayer threads, where more than one human is in the conversation. In her words, people are often just talking to each other and the agent should only jump in when someone actually loops it in. That is the behaviour most chat bots get wrong in the obvious direction, by answering every message that lands in a channel they were invited to.
What was not posted
No accuracy figure, no latency figure, no cost figure. There is no description of what the gate sees, whether it is the single message, the thread, or some window of recent turns, and no mention of what happens on a borderline call. Runkle did not say how often the gate fires or what the false quiet rate looks like. So this is one team's design choice reported by the person who shipped it, not a measured result.
For people building on it
A should-I-reply classifier is about the cheapest thing you can put a model in front of, because the output is one bit and the prompt is a few messages of context. It is also the one with the least visible failure mode. A wrong yes shows up immediately as a bot interrupting two colleagues. A wrong no shows up as nothing. Nobody files a ticket saying the agent stayed silent in a thread where it should have helped, because the people in that thread mostly did not notice there was an agent in the room.
That asymmetry is worth designing around before you ship the gate, not after. If you are copying this pattern, log the negative decisions with the context that produced them, so you have something to review when someone eventually says the bot has been useless lately. Sampling the no calls is the only way to find out whether the gate is well calibrated or whether it has quietly learned to say no to everything, which is the cheapest way for a classifier to look correct.
Runkle's post does not say whether her team does any of that. It says the gate exists and that it helps in multiplayer threads, which is a narrower and more honest claim than most Slack agent posts make.
Know what Jev is good at before you bet on it
One email a day. Under three minutes. What builders found out about Jev yesterday.
