Temporal reveals a serverless option for its Durable Execution platform. Temporal adds a serverless option to Durable Execution, simplifying scalable, crash‑proof long‑running workflows for production and agentic AI. This lowers the operational barrier to running durable orchestrations and maps directly to building resilient agent pipelines (Principle 09).

OpenAI brings GPT-5-level reasoning to its speech models. OpenAI launches realtime speech models with GPT‑5‑class reasoning, 128k context, live translation, and faster streaming transcription. That turns voice into a high‑bandwidth, long‑context channel for agents—changing interaction design, context engineering, and orchestration patterns you must support (Principles 06 & 09).

How Sakana trained a 7B model to orchestrate GPT-5, Claude Sonnet 4 and Gemini 2.5 Pro. Sakana ships an RL Conductor that trains a compact 7B to dynamically coordinate top LLMs, powering cheaper multi‑agent orchestration in Fugu. This demonstrates a practical conductor pattern for building cost‑efficient, learned orchestration layers and highlights the need for graphed context propagation and routing primitives (Principles 09 & 11).

GitHub Improves Token Efficiency in Agentic Workflows. GitHub instruments agentic workflows at the API proxy and applies output constraints to significantly reduce token consumption and recurring CI costs. Those engineering levers are immediately useful for teams running agent fleets—apply proxy‑level observability and output shaping to control cost and preserve throughput (Principles 06 & 11).

Perplexity Launches Mac App with Personal Computer. Perplexity releases a macOS app that hybridizes local file/app access with server inference inside an auditable, sandboxed orchestrator called Personal Computer. The hybrid, sandboxed host is a concrete model for preserving data locality, developer control, and auditability when agents need desktop context—consider this pattern for Gate and Tech‑Island designs (Principles 07, 15 & 06).