July 2026 Summaries

Temporal's Workflow Streams, currently in public preview for Python and TypeScript, is a built-in streaming abstraction designed to provide exactly-once, ordered delivery of a single Workflow's output to subscribers, bypassing the need for a separate relay like a Redis pub/sub system. This system, however, only supports one-way communication through long-polling, meaning it cannot replace a durable session layer which is necessary for capabilities such as real-time push delivery, bidirectional communication, multi-participant visibility, offline delivery, conversation branching, and human handover. While Workflow Streams excels at delivering durable, ordered messages from a Workflow, a durable session layer is required for more complex interaction scenarios where feedback from clients or multiple participant visibility is needed. Users already utilizing a Redis pub/sub relay should evaluate whether Workflow Streams matches their needs, as migrating might not be beneficial if their current setup accommodates bidirectional input, multiple agents, or human operator involvement.

LiveObjects provides a Java-compatible API that manages shared state efficiently using a path-based model, which was initially redefined in JavaScript. This approach allows developers to reference a location in an object tree rather than a specific object, ensuring that references remain valid even when the underlying object changes, thus avoiding stale data and unnecessary rebinding. The API supports type inference and offers flexibility in data handling through constructs like `LiveMap` and `LiveCounter`, enabling robust operations such as tracking leaderboards, maintaining AI session states, or synchronizing configuration panels. This is achieved without the need for a coordination layer, using paths as the primary means of interaction. The path API is integrated with the standard Objects plugin, requiring no additional installation and offering a comprehensive method surface for developers through the LiveObjects Java quickstart and PathObject reference.

The integration of AI into platforms has prompted many companies to reconsider their approach to maintaining realtime infrastructure, as AI conversations demand a shift from signaling to delivery models, which require transmitting responses token-by-token without interruptions. Fin, formerly Intercom, opted to transition from its in-house realtime system, Nexus, to a managed service, achieving higher reliability and reduced operational costs. This decision reflects a broader trend where businesses evaluate the opportunity cost of maintaining self-built systems versus focusing on core product innovations. The complexities and expenses of sustaining realtime infrastructure—estimated at $100K to $200K annually—are compounded when AI-specific functionalities are added, making managed solutions like Ably AI Transport an attractive alternative for ensuring uninterrupted delivery and freeing engineering resources for product development.

AI Transport, developed by Ably, is a durable session layer designed to enhance agent-to-human communication by addressing the limitations of HTTP's request-response model in scenarios where AI agents operate for extended periods. Unlike traditional HTTP, which struggles with connection reliability, bidirectional communication, and state synchronization, AI Transport uses a pub/sub model to decouple agents from clients, allowing for session continuity even when connections drop or devices switch. By inserting a durable session between the agent and user, AI Transport manages connection recovery, enables real-time bidirectional communication, and maintains state continuity independently of backend activities. This approach ensures that both agents and clients can disconnect and reconnect without losing session data, facilitating multi-agent coordination and eliminating the need for custom buffers. AI Transport is designed to integrate seamlessly into existing architectures with minimal code changes, making it a promising addition to the AI agent stack alongside durable memory and execution. While it focuses on human-agent interactions, backend agent-to-agent coordination remains outside its scope, marking a clear boundary for its application.