Launched this week
Humalike
Give your AI agents the social intelligence they're missing
726 followers
Give your AI agents the social intelligence they're missing
726 followers
Today's models are capable enough. Smart enough. Fast enough. But we still feel they don’t fit in the room. Humalike is building the behavioral infrastructure for humanlike AI agents. The social skills & proactiveness your agents have been missing. APIs, models, benchmarks.
This is a really interesting angle — most of the conversation around AI agents focuses on capability (can it use the right tools, in the right order?) but the social layer is almost always missing. I've been studying agentic AI recently and the gap between 'task completed correctly' and 'response that feels right for the situation' is huge. How are you thinking about measuring whether an agent is actually reading the room vs. just following social scripts?
Humalike
@giulia_lemme The reason the current solution is split into various components is because just one didn't solve the problem at all. When you combine all components, agent behaves accordingly! tysm for the supp!
@giulia_lemme Hey, thanks for question! It's hard to evaluate and there is very little pre-existing research on this topic. We try to evaluate different capabilities separately - "was this social" is super hard to answer, but "did this agent follow a norm of behavior, style and humar that this group operates in" is easier. We still need to improve on evaluations side, but one of the steps we did was releasing an open-source benchmark: https://arxiv.org/abs/2606.14600
Documentation.AI
How does Humalike adapt when community norms evolve over weeks or months instead of remaining static?
Humalike
@roopreddy Good question! Norms are extracted from the live transcript, it re-runs extraction on recent windows (new jokes, diff behavior, etc), then Social Memory ingests info continuously. Both combined deliver really well!
@roopreddy That's what inspired us to create Norms API in the first place. People often set up personality once and forget about it. Norms API is meant to be used continously, updating the agent behaviour every few messages.
Is there analytics showing why an agent chose not to respond? That would be incredibly valuable for debugging.
Humalike
@ranjan_kumar45 The component social observability allows you to control that (among extra usefull things)!!
@ranjan_kumar45 Valid point, we already have that for internal debugging but we didn't include it in public API yet.
How are you actually measuring "humanlike" behavior beyond the benchmarks you ship, and can customers plug in their own eval scenarios to test against their specific use case?
@ensarokunakol At the end of the day it's your agent, so you can evaluate it in whatever way you want. We want to allow you to forget about interaction / social part so you can focus on what matters for your agent.
Humalike
@ensarokunakol To be honest, that's not something we have absolutely figured it out yet. We have some things, observability, intuition and "simple" benchmarks, there is still a long way to go upon how to measure. About eval scenarios, yes! The API "Social Observability" does exactly that!!
StartupBase
The social intelligence angle is a sharp wedge. Most agent tooling optimizes for finishing the task and forgets how the interaction actually lands. In practice, are you scoring tone and context, or injecting it into the responses themselves? Feels like something multi-agent setups are going to need soon.
@attacomsian Thanks for the support. We equip your agent with context and judgment it needs to perform better.
Humalike
@attacomsian Totally! One of the components is social observability, which we use ourselves to figure out how to eval scoring. In practice, if we go simple, if you don't complain and overall feel satisfied with the interaction, it means you had a good experience! I think we can all relate to how annoying, impersonal and generally unaware agents are (in any usecase / product) :))
AISA AI Skills Test
the turn-taking problem is so real. I've seen plenty of AI agents that are technically capable but socially exhausting — they jump in too fast, over-explain, and never read the room. curious how you're benchmarking this though — what does 'good' social behavior look like as a metric? is it response timing, or something more nuanced like knowing when a user is thinking vs actually done talking?
@ozandag It's a combination of factors and the ultimate judges are humans: do they speak with him or ignore him? Do they get annoyed by him? Do they trust him? But this can be only evaluated in long time horizon, so we also look at short-term as you noted
Humalike
@ozandag 100%! Experience should just feel right. Our metric isn't fully acc yet, but if you don't complain and have a good experience, that means it's good :))
Really interesting direction. I am curious that for AI chat products, do you think agents will make users feel being manipulated when they can read emotional signals or adapt their tone in real time?
@xinrui1 I think the opposite, users will feel more understood and less annoyed. But manipulation is real risk we have to look out for.
Humalike
@xinrui1 That's a sketchy topic, I think that we as humans use / act with diff behaviors consistently, one of them being manipulation. Humalike pushes in the direction of agents feeling human, with good intention.