←
taras1,003 posts
taras avatar

taras

@voidisomorphism

🇺🇦 building agent-swarm.dev in the open, some thoughts in taras.lol my baby is @agent__swarm

◷ Joined February 2012▧ 150 Photos and videos
taras avatar
taras@voidisomorphism·Oct 16

Introducing Open Oncall: an open source workflow that will investigate, triage and propose fixes for your production incidents. Provider agnostic, because your sleep should not depend on Devin being up. Powered by @agent__swarm.

0:30
taras avatar
taras@voidisomorphism·8h

they have thee source code of word, and still underperform sol

Post attachment
taras avatar
taras@voidisomorphism·8h

Microsoft is the first "lab" that releases a model and the first chart shows its model as the worst one

Post attachment
taras avatar
taras@voidisomorphism·20h

Entropy is where I'd push back tbh. Randomness mostly produces plausible garbage in prod. What compounds is the scaffolding and software: review gates, shared memory, a queue that routes by difficulty. Mixed models, ok. Cheap for drafting, frontier for code and review. To make it real? Use @agent__swarm

Swarms are quickly becoming one of the most important topics in Silicon Valley, sharing a few notes from recent conversations: - The naive view of swarms is that they’re a sum-of-parts method of parallelization / brute force, but we’re empirically seeing the “shape” of swarm intelligence be very different. - The way swarms achieve this is through a combination of parallelization, entropy, and specialization, leading to emergent solutions that aren’t covered by a small number of brilliant frontier models acting as central planners. - It’s also not obvious that the economics of swarms follow higher cost = better output. Swarms that mix and match frontier and non-frontier models with different reasoning levels may be more economical than a small number of frontier agents, and broadly, there should be a new Pareto curve here. - Swarms will be orders of magnitude more token-consumptive than agents, just as agents were compared to humans + models, further driving constraints in inference supply. I think it’s important to build intuition here so have been running some experiments that I’ll share soon. A great way to get started with swarms is the /swarms command in the new google anti-gravity. A few of the smartest thinkers on this topic are @alehandromz @jsnnsa @gokhan_egri @pirroh @amasad Love to hear other POVs on swarms if folks have them!!

Borja Perez avatar
⟳ taras retweeted
Borja Perez@borjaperfra·Nov 17

Artículo completo sobre la nueva arquitectura que hemos probado para coordinar agentes junto a @agent_swarm Deepseek + Clef + reglas superan a Opus 5.5 Todo el experimento, la metodología y la arquitectura completa del sistema de agentes están ahí 👇

Helmcode avatar
⟳ taras retweeted
Helmcode@helmcode·Nov 17

Full write-up on the new agent coordination architecture we tested with @agent_swarm DeepSeek + Clef + rules beat Opus 5.5 The whole experiment, the methodology and the full agent system architecture are all in there 👇 Next steps: more experiments

Borja Perez avatar
⟳ taras retweeted
Borja Perez@borjaperfra·Nov 16

¡Creo que tenemos las 3 charlas para el Café Helmcode BCN! - On building agents that actually learn por @voidisomorphism y @ecura - Aprender más rápido con IA usando ciencia del aprendizaje y convertirlo en contexto para tus agentes por @adoviguera 👀❤️ - Cómo aprenden las máquinas con RL @UbaldoHervas Queda alguna entrada. Apúntate o ayúdame a rularlo. 📆 24 de noviembre 📍 @itnig café 🕧 17:30-21:00 Ha habido muchas charlas guapísimas. Nos ha pasado lo mismo que en el Café Helmcode de Madrid. Ninguna chica ha propuesto charla 😣 Tenemos que conseguir cambiar esto. En el banquillo, por si da tiempo o se nos cae alguien quedan las charlas de @G_Programming, @_MarcosValera o @bernvadell 👀 (o para el siguiente café).

Post attachment
Javi Vázquez avatar
⟳ taras retweeted
Javi Vázquez@xabivaz·Nov 15

Muajaja! Parece que el agent swarm comienza a funcionar autónomo a pleno redimiento!

Post attachment
Helmcode avatar
⟳ taras retweeted
Helmcode@helmcode·Nov 16

👀 DeepSeek just outperformed Claude Opus 5.5 in an AI agent coordination benchmark. We tested a different architecture with our friends at @agent__swarm, adding explicit rules and Clef, a decision model from @Cloudflare, to support agent coordination. The results? DeepSeek V4.1 Flash went from 83% to 100% success on evaluated runs. Claude Opus 5.5 scored 93%. The sample is still small, and the runs weren't identical, so there's more testing to do. But the result is promising: we may be able to improve agent performance by changing the architecture, without always needing a larger model. Huge credit to @voidisomorphism and @ecura for running the experiment. Next, we're expanding the tests to more software engineering tasks. We'll publish a technical article and video breaking down the architecture, methodology and results next days. Full benchmark 👇 api.desplega.agent-swarm.dev…

Post attachment
Borja Perez avatar
⟳ taras retweeted
Borja Perez@borjaperfra·Nov 16

👀😏 Hemos probado una arquitectura nueva para coordinar agentes de IA y los resultados con DeepSeek han superado a Claude Opus 5.5 en el benchmark. La idea es bastante interesante. 👇 En vez de dejar todas las decisiones al LLM, hemos introducido reglas y Clef, un modelo de decisión de @Cloudflare, para ayudar a coordinar el trabajo de los agentes. ¿Los resultados? DeepSeek V4.1 Flash pasa del 83% al 100% de éxito en las ejecuciones evaluadas. Opus se queda en el 93%. El experimento lo han hecho nuestros amigos de @agent__swarm , con quienes estamos colaborando para investigar cómo funcionan los modelos abiertos en sistemas multiagente. Seguidles, que están haciendo cosas muy interesantes. Y también a @voidisomorphism y @ecura La muestra todavía es pequeña y hay bastante que investigar, pero creo que esto abre una vía muy prometedora: mejorar el rendimiento de los agentes cambiando la arquitectura, sin necesidad de utilizar siempre un modelo más grande. Vamos a repetir el experimento con más tareas de ingeniería de software. La semana que viene publicaremos un artículo y vídeo explicando la arquitectura, la metodología y todos los resultados. Os dejo el benchmark por aquí: api.desplega.agent-swarm.dev…

Post attachment
Marce Coll avatar
⟳ taras retweeted
Marce Coll@dzibanisms·Nov 13

deterministic simulation testing for games is fun

0:43
taras avatar
taras@voidisomorphism·Nov 12

Brutal @GuliMoreno, a must read!

Y si no vuelven? Es la crítica más frecuente a @exp8fellowship. Pero la mejor manera de quedarse atrás es retener a los mejores lejos de la frontera. La labor de un pensionado no es volver. Es devolver.