June 2026 Summaries

Stream and Agora both offer robust platforms for real-time communication, video, messaging, and interactive features, but they cater to different needs based on the level of control and integration required by a team. Stream provides high-level application-layer APIs, SDKs, and prebuilt UI components for chat, video, activity feeds, and AI moderation, making it ideal for teams looking to integrate production-ready features quickly without building the underlying infrastructure. In contrast, Agora focuses on media-layer control and conversational AI, providing the infrastructure for audio, video, and interactive media delivery, which is suitable for teams that require fine-grained control over media delivery, such as those building large-scale broadcast systems or AI-driven voice interfaces. While Stream offers a comprehensive platform with integrated moderation, chat, and activity feeds, Agora gives developers the flexibility to build custom solutions by offering separate products for chat and moderation, requiring more development effort but allowing for greater customization at the media layer. The choice between the two platforms ultimately depends on whether a team prioritizes a ready-to-use feature set with integrated components or the ability to customize and control the media and AI layers extensively.

Chat application synchronization involves managing different types of state—durable content, mutable derived state, and ephemeral state—each requiring distinct synchronization strategies. Durable content such as messages and edits must be reliably replayed and survive app reinstallation, while mutable derived states like read receipts necessitate order-sensitive merging to ensure accuracy. Ephemeral states, including presence and typing indicators, are relevant only in real-time and should expire rather than be replayed. Proper synchronization involves first pulling updates from the server to avoid conflicts, followed by replaying queued local writes with idempotent transaction IDs to ensure seamless reconnection. Push notifications should not be used as a transport mechanism for data synchronization but merely as a wake-up signal to initiate a delta sync. The architecture for chat applications should rely on a server-authoritative log for durable content, ensuring that clients can reconcile changes accurately upon reconnection, with mechanisms in place to handle offline storage and ensure retries are safe, particularly in unreliable network conditions.

Large group calls efficiently maintain low latency by using Selective Forwarding Units (SFUs) instead of mesh networks or Multipoint Control Units (MCUs), which helps to avoid bandwidth issues and reduces delays. SFUs route streams without remixing, allowing each client to upload a single stream to a media server, which then forwards only the necessary video quality to each participant, optimizing bandwidth and device performance. WebRTC's real-time transport, short buffers, and adaptive encoding further contribute to keeping delays under a few hundred milliseconds, even in calls with over 50 participants. Media servers are strategically distributed across regions to minimize latency by connecting users to the nearest available server, while client-side limitations are managed through UI strategies like spotlight modes and selective subscription, ensuring only essential video streams are decoded. This approach is supported by SDKs that handle media layers, enabling developers to focus on user interface design rather than complex server management.

In the outlined process, a real-time background removal tool is developed using Vision Agents and Stream Video, capitalizing on models like SAM 2 and YOLO11n to handle person detection and segmentation. The approach involves a Python agent joining a Stream call as a participant to process video frames, allowing for a local preview with a virtual background while keeping the raw video intact for recordings. By utilizing a participant pattern, the solution circumvents the need for complex transport and codec handling, enabling seamless integration with Stream's server SDK. The system's architecture supports configurable settings for background customization and utilizes efficient processing techniques, such as morphological operations and Gaussian blur, to refine segmentation masks for a smooth compositing result. This setup allows for flexible adaptation to other real-time video processing tasks, demonstrating the potential of Vision Agents and Stream Video in enhancing video call experiences with minimal overhead.

Subito, Italy's largest re-commerce marketplace, transformed its buyer-seller messaging system by partnering with Stream, allowing them to scale their infrastructure without expanding their internal team. Faced with the complexity and resource demands of maintaining an in-house messaging system, Subito sought a solution that could handle marketplace-specific needs, such as offer negotiations and sophisticated moderation, while ensuring reliability and scalability across web, iOS, and Android platforms. The decision to collaborate with Stream resulted in a seamless migration of millions of conversations, enhanced fraud prevention, and improved transaction flows, ultimately allowing Subito to focus on its core business of enhancing user trust and transaction efficiency. The partnership with Stream not only optimized Subito's messaging infrastructure but also contributed to platform-wide improvements that benefited other Stream clients, demonstrating the value of strategic collaborations in managing complex technical systems.

In 2025, a major mobile gaming publisher faced the challenge of migrating from the soon-to-be-deprecated Community Sift to Stream's AI moderation stack, which supports over 750 million monthly users and multiple languages. The transition, set within a six-month timeline, required maintaining moderation quality without disrupting established workflows. The publisher's rigorous vendor evaluation process included criteria such as detection quality across diverse languages, low latency under high production loads, and adaptability to the publisher's specific needs. During the Request for Proposal (RFP) process, Stream demonstrated its capabilities through extensive testing and iteration, focusing on improving language detection accuracy and moderating abilities. A critical component of the migration involved shadow testing, enabling both systems to run simultaneously without affecting live operations, allowing for fine-tuning based on discrepancies. The migration process highlighted the importance of dedicated infrastructure, comprehensive vendor testing, and on-site workshops that identified and addressed gaps in moderation tools, ensuring a seamless transition and enhanced player safety.

A vision agent is an AI system capable of processing live video and audio streams in real-time to provide immediate responses, distinguishing it from batch computer vision pipelines and voice agents that only handle static images or audio. Vision agents integrate live video, audio input, a model that understands both, and a response mechanism that operates within the same conversation, all under a second. Building such agents involves complex integration of video transport, speech-to-text, language models, and computer vision, which can be streamlined using the Vision Agents open-source Python framework by Stream. This framework simplifies the development process by managing the technical complexities like data transport, synchronization of various models, and latency management, allowing developers to focus on customizing their vision agents with different models for diverse applications such as telehealth, fitness coaching, and retail, among others. The framework supports integration with numerous platforms and models, offering flexibility for developers to choose between real-time API or custom pipelines for more control. The article highlights the efficiency and potential applications of vision agents across various industries, emphasizing the importance of integrating fast and natural voice capabilities for effective interaction.

The text outlines the process of building an online marketplace platform using Stream's AI agent skills, which include features such as chat, activity feed, moderation, and audio/video integration. These skills enhance developers' productivity by streamlining the integration of complex functionalities, enabling the creation of community-based and enterprise applications at scale. The guide specifically focuses on developing an iOS marketplace app using Stream's four main products, with detailed steps for installation and configuration using Xcode, and highlights the importance of using AI-assisted coding tools and Stream API credentials. It also discusses troubleshooting potential issues that may arise during the integration of chat, feeds, and video features, and emphasizes the flexibility and scalability offered by Stream Skills across various platforms like Android, Swift, React Native, and more. The guide concludes with a demonstration of the app's capabilities, including real-time messaging, video calls, and moderation, and encourages further exploration and extension of the marketplace app using available resources.