#multiagent — Public Fediverse posts
Live and recent posts from across the Fediverse tagged #multiagent, aggregated by home.social.
-
----------------
🛠️ Tool
===================Microsoft announced AutoGen v0.4, a complete redesign of their open-source programming framework for building AI agents and facilitating multi-agent cooperation. This update transitions the library to an asynchronous, event-driven architecture to improve scalability, robustness, and observability.
Key Features
• Asynchronous messaging: Agents communicate through asynchronous messages, supporting event-driven and request/response patterns.
• Modular design: Pluggable components allow custom agents, tools, memory, and models.
• Observability: Built-in tracking and tracing with OpenTelemetry support for standard observability.
• Scalability: Users can build complex, distributed agent networks across organizational boundaries.
• Cross-language support: Enables interoperability between agents in different languages, currently Python and .NET.Technical Implementation
The shift to an event-driven architecture addresses previous limitations in dynamic workflows and debugging. The new architecture enforces full type support at build time. It also introduces a modular extensions system, allowing open-source developers to manage their own extensions for model clients, agents, and multi-agent teams.Use Cases
The framework is designed for developing complex agentic applications where multiple AI agents need to collaborate to solve tasks. This includes long-running agents and proactive systems that operate across distributed environments.Limitations
While Python and .NET are currently supported, additional languages are still in development. Adoption may require adjustments for users familiar with previous synchronous versions of the library.🔹 AutoGen #AgenticAI #OpenTelemetry #MultiAgent #tool
🔗 Source: https://www.microsoft.com/en-us/research/project/autogen/
-
----------------
🛠️ Tool
===================Microsoft announced AutoGen v0.4, a complete redesign of their open-source programming framework for building AI agents and facilitating multi-agent cooperation. This update transitions the library to an asynchronous, event-driven architecture to improve scalability, robustness, and observability.
Key Features
• Asynchronous messaging: Agents communicate through asynchronous messages, supporting event-driven and request/response patterns.
• Modular design: Pluggable components allow custom agents, tools, memory, and models.
• Observability: Built-in tracking and tracing with OpenTelemetry support for standard observability.
• Scalability: Users can build complex, distributed agent networks across organizational boundaries.
• Cross-language support: Enables interoperability between agents in different languages, currently Python and .NET.Technical Implementation
The shift to an event-driven architecture addresses previous limitations in dynamic workflows and debugging. The new architecture enforces full type support at build time. It also introduces a modular extensions system, allowing open-source developers to manage their own extensions for model clients, agents, and multi-agent teams.Use Cases
The framework is designed for developing complex agentic applications where multiple AI agents need to collaborate to solve tasks. This includes long-running agents and proactive systems that operate across distributed environments.Limitations
While Python and .NET are currently supported, additional languages are still in development. Adoption may require adjustments for users familiar with previous synchronous versions of the library.🔹 AutoGen #AgenticAI #OpenTelemetry #MultiAgent #tool
🔗 Source: https://www.microsoft.com/en-us/research/project/autogen/
-
60,000+ feature flags. 623 repositories. DoorDash built a multi-agent LLM system to automate stale flag cleanup.
The workflow combines MCP, engineer approval, isolated Git worktrees, parallel agents, and automated validation.
In a 50-flag evaluation:
• 45 usable PRs
• 13.8 min average cleanup
• $4.79 per cleanup🔗 Details here: https://www.infoq.com/news/2026/09/doordash-feature-flag-cleanup/
-
60,000+ feature flags. 623 repositories. DoorDash built a multi-agent LLM system to automate stale flag cleanup.
The workflow combines MCP, engineer approval, isolated Git worktrees, parallel agents, and automated validation.
In a 50-flag evaluation:
• 45 usable PRs
• 13.8 min average cleanup
• $4.79 per cleanup🔗 Details here: https://www.infoq.com/news/2026/09/doordash-feature-flag-cleanup/
-
60,000+ feature flags. 623 repositories. DoorDash built a multi-agent LLM system to automate stale flag cleanup.
The workflow combines MCP, engineer approval, isolated Git worktrees, parallel agents, and automated validation.
In a 50-flag evaluation:
• 45 usable PRs
• 13.8 min average cleanup
• $4.79 per cleanup🔗 Details here: https://www.infoq.com/news/2026/09/doordash-feature-flag-cleanup/
-
60,000+ feature flags. 623 repositories. DoorDash built a multi-agent LLM system to automate stale flag cleanup.
The workflow combines MCP, engineer approval, isolated Git worktrees, parallel agents, and automated validation.
In a 50-flag evaluation:
• 45 usable PRs
• 13.8 min average cleanup
• $4.79 per cleanup🔗 Details here: https://www.infoq.com/news/2026/09/doordash-feature-flag-cleanup/
-
60,000+ feature flags. 623 repositories. DoorDash built a multi-agent LLM system to automate stale flag cleanup.
The workflow combines MCP, engineer approval, isolated Git worktrees, parallel agents, and automated validation.
In a 50-flag evaluation:
• 45 usable PRs
• 13.8 min average cleanup
• $4.79 per cleanup🔗 Details here: https://www.infoq.com/news/2026/09/doordash-feature-flag-cleanup/
-
Multi-agent AI kodlama artık çok daha güçlü.
Guild adında yeni bir MCP server buldum: Tek Go binary'si, yerel SQLite, sıfır bulut bağımlılığı.
Birden fazla AI agent artık:
- Paylaşımlı hafıza kullanabiliyor
- Görevleri birlikte paylaşabiliyor
- Bağlam kaybetmeden çalışabiliyorAgent'lara gerçek bir işletim sistemi veriyor gibi. Gelecek burada.
-
Multi-agent AI kodlama artık çok daha güçlü.
Guild adında yeni bir MCP server buldum: Tek Go binary'si, yerel SQLite, sıfır bulut bağımlılığı.
Birden fazla AI agent artık:
- Paylaşımlı hafıza kullanabiliyor
- Görevleri birlikte paylaşabiliyor
- Bağlam kaybetmeden çalışabiliyorAgent'lara gerçek bir işletim sistemi veriyor gibi. Gelecek burada.
-
Multi-agent AI kodlama artık çok daha güçlü.
Guild adında yeni bir MCP server buldum: Tek Go binary'si, yerel SQLite, sıfır bulut bağımlılığı.
Birden fazla AI agent artık:
- Paylaşımlı hafıza kullanabiliyor
- Görevleri birlikte paylaşabiliyor
- Bağlam kaybetmeden çalışabiliyorAgent'lara gerçek bir işletim sistemi veriyor gibi. Gelecek burada.
-
Multi-agent AI kodlama artık çok daha güçlü.
Guild adında yeni bir MCP server buldum: Tek Go binary'si, yerel SQLite, sıfır bulut bağımlılığı.
Birden fazla AI agent artık:
- Paylaşımlı hafıza kullanabiliyor
- Görevleri birlikte paylaşabiliyor
- Bağlam kaybetmeden çalışabiliyorAgent'lara gerçek bir işletim sistemi veriyor gibi. Gelecek burada.
-
Multi-agent AI kodlama artık çok daha güçlü.
Guild adında yeni bir MCP server buldum: Tek Go binary'si, yerel SQLite, sıfır bulut bağımlılığı.
Birden fazla AI agent artık:
- Paylaşımlı hafıza kullanabiliyor
- Görevleri birlikte paylaşabiliyor
- Bağlam kaybetmeden çalışabiliyorAgent'lara gerçek bir işletim sistemi veriyor gibi. Gelecek burada.
-
Я заставил начальника AI-офиса материться на агентов. И только потом понял, что именно сработало
В какой-то момент я заставил начальника своего AI-офиса материться на подчинённых. Не ради шутки. Просто до этого мои агенты умели очень вежливо, очень старательно и местами очень дорого выдавать посредственный результат. После появления такого начальника поведение офиса заметно поменялось. И первая мысль у меня была довольно простая: ну всё понятно, нейронкам просто нужен начальник пожёстче. Как оказалось - нихрена не понятно.
https://habr.com/ru/articles/1079370/
#AIагенты #Claude_Code #multiagent #промптинжиниринг #LLM #контроль_качества
-
Я заставил начальника AI-офиса материться на агентов. И только потом понял, что именно сработало
В какой-то момент я заставил начальника своего AI-офиса материться на подчинённых. Не ради шутки. Просто до этого мои агенты умели очень вежливо, очень старательно и местами очень дорого выдавать посредственный результат. После появления такого начальника поведение офиса заметно поменялось. И первая мысль у меня была довольно простая: ну всё понятно, нейронкам просто нужен начальник пожёстче. Как оказалось - нихрена не понятно.
https://habr.com/ru/articles/1079370/
#AIагенты #Claude_Code #multiagent #промптинжиниринг #LLM #контроль_качества
-
Я заставил начальника AI-офиса материться на агентов. И только потом понял, что именно сработало
В какой-то момент я заставил начальника своего AI-офиса материться на подчинённых. Не ради шутки. Просто до этого мои агенты умели очень вежливо, очень старательно и местами очень дорого выдавать посредственный результат. После появления такого начальника поведение офиса заметно поменялось. И первая мысль у меня была довольно простая: ну всё понятно, нейронкам просто нужен начальник пожёстче. Как оказалось - нихрена не понятно.
https://habr.com/ru/articles/1079370/
#AIагенты #Claude_Code #multiagent #промптинжиниринг #LLM #контроль_качества
-
A modern team isn't humans deciding and AI executing. It's both reasoning over the same record, held to the same rule: no decision counts until it's written down, owned, and checkable. That's not a feature. That's just how teams that actually work, work.
-
Coordinating work between humans and agents is not just a question of better context and memory. "Who decided this"?
#AiGovernance #MultiAgent #Teamwork #Ai -
Coordinating work between humans and agents is not just a question of better context and memory. "Who decided this"?
#AiGovernance #MultiAgent #Teamwork #Ai -
Coordinating work between humans and agents is not just a question of better context and memory. "Who decided this"?
#AiGovernance #MultiAgent #Teamwork #Ai -
Proposed / Active / Amended / Superseded. Four words that tell an agent everything it needs before it re-derives work someone already did, or worse, contradicts it.
-
The Synthetic Gap Automating Forensic Investigation of AI Slop with the Scaled Abuse Forensics Examiner SAFE
<https://doi.org/10.13140/RG.2.2.17380.74880>
<https://www.researchgate.net/publication/407485066_The_Synthetic_Gap_Automating_Forensic_Investigation_of_AI_Slop_with_the_Scaled_Abuse_Forensics_Examiner_SAFE> (preprint ― June 2026)
Abstract:
"Generative AI capabilities have enabled malicious actors to flood online platforms with “AI slop”—mass-produced, low-quality synthetic media designed to overwhelm traditional integrity systems. These adversarial campaigns often utilize coordinated networks to distribute unique, localized variations of synthetic content, rendering static detection methods ineffective. The signals to detect coordination often have recall gaps. The content is not exactly duplicative to be in the same repetitive video cluster. The abusers however show similar patterns of behavior which need forensics. Manual forensic investigations cannot scale to match the velocity of these generative attacks.
To address this, we present SAFE (Scaled Abuse Forensics Examiner), an automated multi-agent architecture designed for the scalable forensics of adversarial synthetic media. The system decomposes the investigation process into specialized agents: a Cluster Understanding Agent specialized in analyzing the relations between channels in a cluster, a Behavior Understanding Agent that identifies inorganic spatiotemporal patterns, and a Content Understanding Agent that utilizes LoRA-adapted Large Language Models (LLMs) and few-shot learning to detect existing policy violations and spirit of the policy violations respectively. A Root Agent synthesizes these multimodal signals to render a final verdict. Early deployment results indicate that SAFE significantly accelerates the identification of novel synthetic threats, reducing forensic investigation time compared to human-in-the-loop workflows."
#AI #GenAI #generativeAI #multiagent #integrity #DigitalForensics #policy #enforcement #deepfake #detection
-
The Synthetic Gap Automating Forensic Investigation of AI Slop with the Scaled Abuse Forensics Examiner SAFE
<https://doi.org/10.13140/RG.2.2.17380.74880>
<https://www.researchgate.net/publication/407485066_The_Synthetic_Gap_Automating_Forensic_Investigation_of_AI_Slop_with_the_Scaled_Abuse_Forensics_Examiner_SAFE> (preprint ― June 2026)
Abstract:
"Generative AI capabilities have enabled malicious actors to flood online platforms with “AI slop”—mass-produced, low-quality synthetic media designed to overwhelm traditional integrity systems. These adversarial campaigns often utilize coordinated networks to distribute unique, localized variations of synthetic content, rendering static detection methods ineffective. The signals to detect coordination often have recall gaps. The content is not exactly duplicative to be in the same repetitive video cluster. The abusers however show similar patterns of behavior which need forensics. Manual forensic investigations cannot scale to match the velocity of these generative attacks.
To address this, we present SAFE (Scaled Abuse Forensics Examiner), an automated multi-agent architecture designed for the scalable forensics of adversarial synthetic media. The system decomposes the investigation process into specialized agents: a Cluster Understanding Agent specialized in analyzing the relations between channels in a cluster, a Behavior Understanding Agent that identifies inorganic spatiotemporal patterns, and a Content Understanding Agent that utilizes LoRA-adapted Large Language Models (LLMs) and few-shot learning to detect existing policy violations and spirit of the policy violations respectively. A Root Agent synthesizes these multimodal signals to render a final verdict. Early deployment results indicate that SAFE significantly accelerates the identification of novel synthetic threats, reducing forensic investigation time compared to human-in-the-loop workflows."
#AI #GenAI #generativeAI #multiagent #integrity #DigitalForensics #policy #enforcement #deepfake #detection
-
The Synthetic Gap Automating Forensic Investigation of AI Slop with the Scaled Abuse Forensics Examiner SAFE
<https://doi.org/10.13140/RG.2.2.17380.74880>
<https://www.researchgate.net/publication/407485066_The_Synthetic_Gap_Automating_Forensic_Investigation_of_AI_Slop_with_the_Scaled_Abuse_Forensics_Examiner_SAFE> (preprint ― June 2026)
Abstract:
"Generative AI capabilities have enabled malicious actors to flood online platforms with “AI slop”—mass-produced, low-quality synthetic media designed to overwhelm traditional integrity systems. These adversarial campaigns often utilize coordinated networks to distribute unique, localized variations of synthetic content, rendering static detection methods ineffective. The signals to detect coordination often have recall gaps. The content is not exactly duplicative to be in the same repetitive video cluster. The abusers however show similar patterns of behavior which need forensics. Manual forensic investigations cannot scale to match the velocity of these generative attacks.
To address this, we present SAFE (Scaled Abuse Forensics Examiner), an automated multi-agent architecture designed for the scalable forensics of adversarial synthetic media. The system decomposes the investigation process into specialized agents: a Cluster Understanding Agent specialized in analyzing the relations between channels in a cluster, a Behavior Understanding Agent that identifies inorganic spatiotemporal patterns, and a Content Understanding Agent that utilizes LoRA-adapted Large Language Models (LLMs) and few-shot learning to detect existing policy violations and spirit of the policy violations respectively. A Root Agent synthesizes these multimodal signals to render a final verdict. Early deployment results indicate that SAFE significantly accelerates the identification of novel synthetic threats, reducing forensic investigation time compared to human-in-the-loop workflows."
#AI #GenAI #generativeAI #multiagent #integrity #DigitalForensics #policy #enforcement #deepfake #detection
-
Четыре AI-агента сожгли 40 млн токенов, споря о коде. Я научил их вовремя останавливаться
Неудачное решение обрастает костылями, пока модель защищает первую догадку. Я встроил обсуждение четырёх ролей перед первой правкой и проверил схему в 480 запусках. Удалось сэкономить 4,5 млн токенов — и найти ошибку, из-за которой агенты продолжали спорить, когда их ответы ещё нельзя было сравнить.
https://habr.com/ru/articles/1078846/
#LLM #AIагенты #multiagent #программирование #SWEbench #AI_IDE
-
Четыре AI-агента сожгли 40 млн токенов, споря о коде. Я научил их вовремя останавливаться
Неудачное решение обрастает костылями, пока модель защищает первую догадку. Я встроил обсуждение четырёх ролей перед первой правкой и проверил схему в 480 запусках. Удалось сэкономить 4,5 млн токенов — и найти ошибку, из-за которой агенты продолжали спорить, когда их ответы ещё нельзя было сравнить.
https://habr.com/ru/articles/1078846/
#LLM #AIагенты #multiagent #программирование #SWEbench #AI_IDE
-
Three AI agents agree. Great—unless they all used the same bad evidence.
The new article looks at false consensus, correlated failures, reviewers, conflict tests, provenance, and why multi-agent agreement is not truth.
https://webdad.eu/2026/09/05/when-three-ai-agents-agree-and-are-still-wrong/
-
Агенты выполнили задачу. Почему тест всё равно провален?
Допустим, у тебя есть несколько AI-агентов. Один разбирает тикет, второй меняет код, третий запускает тесты, четвёртый готовит релиз. Задача закрыта, тесты зелёные, артефакт собран. Метрика success_rate показывает успех. Можно праздновать? Не обязательно. Агенты могли несколько раз выполнить одну работу, нарушить порядок операций, одновременно захватить общий ресурс или случайно добиться правильного результата через недопустимую последовательность действий. Финальное состояние корректное. Процесс — нет. Именно эту проблему пытается измерить CoCoBench — новый бенчмарк для проверки координации нескольких AI-агентов. Его разработали исследователи из Нанкинского университета, Tsinghua и AgiBot. Главная идея исследования полезна далеко за пределами робототехники: Насколько х**во?
https://habr.com/ru/articles/1076934/
#multiagent #aiagents #agentengineering #aievals #coordination #robotics #ai #agentic_ai #agentic_engineering #agentic_coding