I built Aisir because I often found myself wanting multiple AI perspectives for complex coding/analysis tasks and was kind of frustrated by the blindspots of single models. So then the natural thing to try was, what if rather than just queueing each model in one go they actually talk to each other and work together to solve the problem.
Aisir uses a "council" approach where agents (currently Gemini, Claude Sonnet, o4-mini, WebSearcher (Gemini based)) deliberate on a query over a number of rounds. A moderator agent guides the discussion, chooses what to do next, identifies issues, and pushes for refinement before a final answer is synthesized. This is obviously overkill on simple queries but I would speculate (obviously biased) that it could beat price per token on the some very complicated queries vs like o1-pro as multiple models working together have a quicker way in token terms to find the right answer than a single model. This is based on my anecdotal experiments vs singular models.
So to explain, the difficult (and very limited) benchmarking I've done on it is using the epoch.ai FrontierMath examples to make sure the result would atleast not be worse than just using the best singular model but turns out it's more likely to answer correctly than any single model. This is slightly obvious in the sense that if a model can't answer one specific question, another one might be able to instead even if they don't talk to eachother. The next test would be to see if there are specific problems that none of the best models can answer individually but can be solved using this council method. Let me know if you find any by hand.
You can see the 'thinking' process unfold, showing each agent's contribution and the moderator's comments.
This is very much an experiment and currently free to use. Running these models is expensive over time, so if it gets traction, I might need to add limits/subscriptions later, but for now, I'm focused on seeing if the core idea is useful.
I'd love to get your feedback, especially on: a) Does the multi-agent approach yield better results for you? b) What kinds of complex problems would this be most useful for? c) Any suggestions for improving? There's a lot of optimization I see with token spent etc but I want to see if this is interesting or valuable to anyone other than me.
I'm esamust, the creator behind Aisir.
I built Aisir because I often found myself wanting multiple AI perspectives for complex coding/analysis tasks and was kind of frustrated by the blindspots of single models. So then the natural thing to try was, what if rather than just queueing each model in one go they actually talk to each other and work together to solve the problem.
Aisir uses a "council" approach where agents (currently Gemini, Claude Sonnet, o4-mini, WebSearcher (Gemini based)) deliberate on a query over a number of rounds. A moderator agent guides the discussion, chooses what to do next, identifies issues, and pushes for refinement before a final answer is synthesized. This is obviously overkill on simple queries but I would speculate (obviously biased) that it could beat price per token on the some very complicated queries vs like o1-pro as multiple models working together have a quicker way in token terms to find the right answer than a single model. This is based on my anecdotal experiments vs singular models.
So to explain, the difficult (and very limited) benchmarking I've done on it is using the epoch.ai FrontierMath examples to make sure the result would atleast not be worse than just using the best singular model but turns out it's more likely to answer correctly than any single model. This is slightly obvious in the sense that if a model can't answer one specific question, another one might be able to instead even if they don't talk to eachother. The next test would be to see if there are specific problems that none of the best models can answer individually but can be solved using this council method. Let me know if you find any by hand.
You can see the 'thinking' process unfold, showing each agent's contribution and the moderator's comments.
This is very much an experiment and currently free to use. Running these models is expensive over time, so if it gets traction, I might need to add limits/subscriptions later, but for now, I'm focused on seeing if the core idea is useful. I'd love to get your feedback, especially on: a) Does the multi-agent approach yield better results for you? b) What kinds of complex problems would this be most useful for? c) Any suggestions for improving? There's a lot of optimization I see with token spent etc but I want to see if this is interesting or valuable to anyone other than me.
Link to the tool: https://aisirai.com