every company needs a cassandra • Solving the decision problem skip to content

Close

On this page<br>cassandra<br>the economics of being annoying<br>knowing when to shut up<br>give her a different brain<br>this could also go badly<br>maybe agents aren’t just workers<br>every company needs a cassandra<br>9 August 2026 an ai agent for the socially expensive work of organizational dissent

every company has had this person. they’re… annoying. when everyone is excited about something, they’ll tell you why it won’t work. when someone drops a competitor’s launch into slack and everyone starts making fun of it, they’ll point out the thing the competitor is actually doing well. and ffs, the annoying part is that they’re often right.

I’ve worked with a few people like this. sometimes I’ve been this person. they can sound like pessimists, but mostly they’re willing to say something everyone else would rather not hear.

organizations have a weird relationship with these people. we say we want disagreement, but every disagreement costs the person making it a little social/political capital. do it often enough and you become the negative person. people stop evaluating the argument and start evaluating the source. “oh, that’s just alice being alice.” eventually alice learns to keep quiet, or stops getting invited to the conversations where everyone wants to move quickly. a lot of times they just leave.

that’s bad, because organizations are terrible at being self-critical without people like alice around. consensus feels good, especially when the people agreeing with you are smart people you respect. a group of intelligent people can build a very convincing explanation for why the thing they already want to believe is true. so I’ve been wondering: what if alice was an ai agent?

cassandra

I’m going to call this agent cassandra, after the woman in greek mythology who could see the future but was cursed so nobody would believe her. cassandra sits in your slack channels like any other work agent. nobody really asks her to do anything. most of the time she just watches.

someone posts a competitor’s new product and everyone piles on about how stupid the architecture is. cassandra might say:

the architecture criticism seems fair, but I think we’re dismissing something important. they’ve reduced the setup process to about five minutes. ours still requires understanding three platform concepts. customers might care considerably more about that than the architectural problem we’re laughing at.

or perhaps you’ve spent a few days discussing the risks of shipping something:

we’ve spent three days discussing the risks of shipping this. as far as I can tell they’re all reversible. I think caution might now be costing us more than the risks we’re avoiding.

(that last one is a good one imo).

cassandra should disagree with pessimists too. “you are a devil’s advocate, disagree with everything” is probably the worst possible prompt for this agent. within a week everyone would ignore her. she’d become the slack equivalent of a smoke detector that goes off whenever somebody makes toast.

her actual job is to form her own view of what is happening, then speak when it differs enough from the consensus to be worth the interruption. sometimes cassandra should look hard for an objection and say nothing. sometimes she should say, “I tried to find a strong counterargument here and couldn’t.” I’d trust her much more after that.

the economics of being annoying

an ai is oddly well suited to this job. cassandra doesn’t want a promotion and doesn’t need the people in the channel to like her. disagreeing with the vp three times this month won’t show up in a performance review. she doesn’t have an ego invested in being right either. if she objects to a project in january and the data proves her wrong in march, she can just say so. being wrong should make her model of the company better.

she also isn’t in the critical path of anything. the human version of this character is hard to work with when you need them to finish the database migration by friday and they’re still challenging every decision around it. at some point the practical cost of the disagreement overwhelms its value. an agent doesn’t have that problem, though it creates a new one: cassandra can be infinitely annoying. so the hard part of building her isn’t so much the prompt, but “deciding” when she gets to speak.

knowing when to shut up

almost all messages should get no response. cassandra could (should?) read everything and maintain a private model of the organization: assumptions, decisions, predictions, rejected alternatives, the customer complaints that keep coming back. then she’d have to decide whether any of it is worth interrupting humans about.

the rough calculation is something like:

importance × disagreement × evidence × novelty

if you’re choosing the color of a button, cassandra can go away. if twenty people are converging on a decision that will take six months to reverse, and she has...