Open AI expanded the capabilities of its Codex software engineering agent that launched last month and also making it available to more users. It’s also making more tools available to developers of voice agents through the OpenAI Agents software development kit. Codex is now able to connect to the web to install dependencies, run tests that require external resources, upgrade packages and more tasks that demand internet connectivity. However, the company said internet access remains switched off by default, and must be enabled for specific environments. Users will be able to control which domains it can access. The capability is coming to ChatGPT Plus, Pro and Teams users first, with Enterprise users expected to get it later. There are several other updates too. For instance, Codex now has the ability to update pull requests when performing a follow up task, and users will also be able to dictate tasks to it, instead of typing them out. Elsewhere, it gets support for binary files when applying patches, improved error messages for setup scripts, and increased limits on task diffs, up from 1 megabyte to 5 megabytes, and higher limits of 10 minutes on setup script durations, up from five minutes previously. OpenAI has also re-enabled live activities for iOS Codex users after fixing a problem pertaining to missed notifications, and removed the two-factor authentication requirement for users with single sign-on or social logins. OpenAI Agents SDK now supports TypeScript, with features including guardrails, handoffs, tracing and the Model Context Protocol, which provides a standardized way for agents to use third-party software tools such as browsers. The Agents SDK launched in March, providing tools for integrating AI models and agents with internal systems. Other updates in the Agents SDK include support for human-in-the-loop approvals, so developers can pause tool executions, serialize and store agent states and approve or reject specific calls, the company said. The Traces dashboard is getting support for Realtime application programming interface sessions, making it easier for developers to visualize voice agent runs. Finally, there’s an updated speech-to-speech model available in the SDK that delivers improved instruction following capabilities, interruption behavior and tool-calling consistency. Developers will also be able to control how fast the voice speaks.