engblogs

A technical guide to Duolingo's production-ready agent platform that decouples agent definition from execution, enables reusable agents via a registry, and delivers durable orchestration, observability, and evaluation through Temporal and the OpenAI Agents SDK.

Duolingo demonstrates a shared agent platform that decouples agent definition from execution to deliver production-ready AI agents with durable orchestration, observability, and cross-surface deployment via Temporal.

Circles harnesses the OpenAI API to power an AI-native telco stack, featuring the AI Concierge and CareX multi-agent architecture, for real-time personalization and autonomous resolution, delivering a 22% ARPU uplift and 9% churn reduction in Singapore (65% autonomous resolution).

Six months to build GPT-Live: a low-latency, full-duplex voice AI system that streams audio, separates media from application logic, and delegates deeper reasoning to frontier models to deliver live, natural conversations in ChatGPT Voice.

Technical analysis of Apple vs. OpenAI: miscommunications, access-control failures, and the evolving narrative around trade secrets and confidential information shaping modern tech litigation.

OpenAI-powered AI Concierge and its CareX multi-agent architecture drive real-time telco personalization, delivering proactive customer engagement that boosts ARPU, reduces churn, and speeds engineering with Codex in Circles’ AI-native stack.

Circles builds an AI-native telco stack powered by OpenAI to deliver real-time, proactive personalization via an AI Concierge and CareX multi-agent architecture, achieving autonomous resolution and ARPU growth.

Six-month, end-to-end rollout of a streaming, full-duplex voice AI system (GPT-Live) that removes turn detectors, integrates frontier models, and delivers sub-second, live conversations at scale through stateful inference, asynchronous delegation, and optimized WebRTC transport.

Six months to a scalable, real-time voice AI: a streaming, full-duplex system that replaces turn detectors with stateful inference and asynchronous delegation to frontier models (GPT-5.5) atop a low-latency WebRTC pipeline with WARP and Instant Connect for live, natural conversations.

A technical critique of Apple's legal dispute highlighting how miscommunication, inadequate offboarding, and weak access controls can blur the line between trade secrets and confidential information.

A technical case study on information governance, access-control failures, and miscommunication in a high-stakes corporate dispute, illustrated by the Apple–OpenAI litigation.

Designing realistic, domain-specific enterprise environments that model complex workflows, enable grounded tool use and stateful evaluation, and advance reinforcement learning for training agents beyond final answers.