OpenAI has introduced the Agents API in public beta, providing developers with a managed harness and infrastructure for building and running long-running, production-ready AI agents. This API leverages the same robust harness and infrastructure that powers Codex and ChatGPT for Work, designed to manage context, efficiently use tools, and coordinate subagents.

The Agents API allows for the creation of agents with a single API call, specifying the model (e.g., GPT-6 Astra), tools, and environment. Developers can choose between OpenAI-hosted sandboxes or integrate with partner environments like Modal, Cloudflare, and Oracle, offering flexible compute, storage, and deployment options.

Key features include advanced context management for long sessions, which automatically compacts earlier context to preserve relevant information, and efficient tool use through tool search and programmatic tool calling. The API also supports multi-agent capabilities, enabling complex tasks to be broken down and delegated to parallel subagents, significantly speeding up workflows.

The underlying Codex harness is open-source, offering transparency into its core logic, while OpenAI handles its operation and continuous improvement. The Agents API is available without additional fees, with costs based on token and tool usage, aiming to empower developers to build sophisticated AI agents for diverse applications.

👉 Read the full article on OpenAI

With the introduction of more capable models like GPT-6 Astra, OpenAI emphasizes the need for developers to rethink and refine their existing agent instructions, including skills, AGENTS.md files, and task prompts. Older practices of extensive handholding and scaffolding are no longer necessary, and can even hinder performance due to bloated context.

The article advises making skill descriptions as short and clear as possible, focusing on when a skill applies rather than over-emphasizing its use. Overly long or contradictory descriptions can lead the model to shorten them, making skill selection less effective. Progressive disclosure for skills with multiple workflows is recommended, using a minimal root document to point to supporting documentation.

For AGENTS.md, developers should frequently revisit instructions to ensure they are still needed, avoiding directives that force the model to read entire project documentation before every edit. GPT-6 Astra is capable of determining what it needs to read contextually. The model's improved ability to run tests and check its own work means previous instructions for these actions can now lead to unnecessary steps.

Developers should also pay attention to how they define decision boundaries and persistence. While previous models might have required strong language to prevent overreach, GPT-6 Astra has better judgment and will not perform unsafe tasks. Therefore, overly cautious language might cause Astra to stop work prematurely. Defining task completion clearly from the start can help Astra persist until the work is fully done.

👉 Read the full article on OpenAI Developers

This article discusses how the introduction of new AI features in the Apple Watch is contributing to the normalization of "always-listening" technology. The integration of advanced voice-activated capabilities and proactive assistance, powered by Generative AI and AI Agents, means the device is constantly processing ambient audio to anticipate user needs and respond to commands. This continuous monitoring, while enhancing convenience, raises significant privacy concerns among users and experts.

The article likely explores the trade-offs between enhanced user experience and potential privacy erosion. It suggests that as such features become commonplace, users may grow accustomed to the idea of their personal devices perpetually monitoring their environment, potentially lowering their guard regarding data collection and usage. The discussion would involve the ethical implications of pervasive listening and the need for clear transparency from tech companies like Apple regarding how audio data is collected, processed, and secured.

Furthermore, the piece probably touches upon the technical aspects of how these AI systems operate on-device versus in the cloud, and the implications for data security. It would highlight the societal shift where constant technological surveillance, once a niche concern, becomes an accepted part of daily life due to the perceived benefits of AI-driven assistance.

👉 Read the full article on TechCrunch