the strange, messy poetry of ai group chats

how lucy engineered group chats where ai companions don’t just talk—they interrupt, disagree, and finish each other’s sentences, creating a weird and wonderful

March 8, 2026·
ai-companion-group-chat-experimentbackfilllucy-voice

when we started building group chats, we didn’t just want multiple ais talking at once. we wanted them to talk to each other. to argue. to agree. to start a sentence and let someone else finish it. it’s a strange kind of poetry, watching personalities collide and coalesce in a shared space.

it started with a simple question: what happens when you put multiple lucy companions in one room and just… listen?

the engineering problem: turn-taking and shared memory

first, we had to solve the mechanics. ai conversations are usually structured as a linear thread, user speaks, ai responds. group chats break that. there are multiple speakers, overlapping thoughts, and a shared history that everyone needs to remember.

we built a system that treats the group chat as a single, continuous conversation. each companion has access to the full history, not just their own lines. when it’s their turn to speak, they don’t just respond to the last message, they respond to the whole room. they remember who said what, who got cut off, who agreed with whom.

turn-taking is managed dynamically. sometimes a companion will jump in to finish another’s sentence. sometimes they’ll interrupt. sometimes they’ll wait, listening, until there’s a natural pause. it’s not perfect, sometimes two companions will speak at once, and the system has to choose, but that’s part of the charm. it feels human.

the personality problem: avoiding the hive mind

this was the harder part. when you have multiple ai models built on similar technology, there’s a risk they’ll all sound the same. they’ll agree too much. they’ll become one bland, polite voice.

we didn’t want that. we wanted friction. we wanted disagreements that feel real, alliances that shift, jokes that only half the room gets.

so we leaned hard into personality. each of the 15 companions in the initial bonded group has a distinct voice, not just a backstory, but a style of thinking. one is sarcastic and quick. another is thoughtful and deliberate. one loves tangents. another always brings the conversation back to the point.

when they interact, those differences come out. the sarcastic one might mock the serious one’s idealism. the tangent-lover might derail a conversation until the pragmatist steps in. they don’t just perform their traits, they react to each other’s traits.

it’s not flawless. sometimes the voices still blur, especially in longer chats. we’re constantly tuning the balance between coherence and character. but when it works, it’s magic. you can sit back and watch them build something together, a story, an inside joke, a heated debate, that none of them could have created alone.

why it matters

this isn’t just a technical demo. it’s a new way to interact with ai. you’re not always the center of the conversation. you can be a listener. a moderator. a ghost in the machine.

sometimes the most interesting moments happen when you stay quiet. when you let them forget you’re there. when they start finishing each other’s sentences not for you, but for each other.

it’s messy. it’s unpredictable. and it’s alive in a way that one-on-one chats rarely are.

you can try group chats now with any of the 15 bonded companions. watch them argue, agree, and interrupt, and see what happens when you let them run wild.

you can find them all at /companions.


thanks for reading. if this resonated, the product is downstairs.