The 2-minute brief for AI builders.

🤖 OpenAI opens managed cloud agents in beta

Want the agent without babysitting the plumbing? OpenAI’s Agents API is now in public beta, offering cloud agents that run with the Codex harness. Read the original post on X →

• OpenAI handles orchestration, long-running sessions, and context management.
• You focus on the parts that differentiate your agent.
• OpenAI runs the agent loop across model calls, tool use, and context, while you control capabilities and choose where code runs and files are handled.
• OpenAI also offers hosted sandboxes where agents can run code, work with files, and produce artifacts.

What it means: Prototype agents that need sustained context or tool use faster, then spend your engineering time on permissions, workflows, and product behavior instead of agent-loop infrastructure.

🎙️ GPT-Live-1 brings fluid voice agents to the API

Voice agents can now keep up with an actual conversation. GPT-Live-1 is available in the API for apps that need listening and speaking to happen at the same time. Read the original post on X →

• The model is built for natural back-and-forth and can work with the models and harness you choose.
• It distinguishes speech from background noise, so ambient sound does not have to stop a conversation.
• Users can interrupt, add details, or change direction before the model finishes responding.
• You can set tone, pacing, expressiveness, language, and response length for the voice experience.

What it means: Test voice support, intake, or concierge flows where customers interrupt naturally, while your backend model continues handling reasoning and tool calls.

🧑‍💻 Cursor Projects keeps coding agents on the job

One task, one chat, one forgotten thread? Cursor’s new Projects feature replaces that pattern with a persistent coordinator agent that manages work over time. Read the original post on X →

• The coordinator agent works in one ongoing thread rather than a fresh chat for every task.
• It can proactively manage work with subagents and improve over time.
• Projects can set reminders, run scheduled tasks, follow pull requests to fix CI issues, and watch Slack for bug reports.
• Agents share memory and generated artifacts, including plans and demos, which sync across devices and agent computers.

What it means: Use a Project for your product’s standing maintenance queue, so CI failures, Slack reports, and follow-ups enter one shared agent workspace instead of disappearing into chat history.

AI Tools I'm Actually Using

These are affiliate links — if you sign up, I may earn a commission at no extra cost to you. I only recommend tools I use myself.

⏰ Replit schedules AI work without constant spend

Your recurring ops work does not need a model call every minute. Replit introduced Routines, scheduled automations that start with deterministic code and call Agent only when reasoning is needed. Read the original post on X →

• Schedule a Routine hourly, daily, or weekly.
• The workflow is designed for recurring work while avoiding unnecessary token use.
• One example collects and organizes sales opportunities, then uses Agent to prioritize the leads needing attention.
• Core and Pro customers can set a budget for each Routine to keep token spend within a defined limit.

What it means: Turn repeatable lead triage, feedback analysis, or reporting into scheduled workflows, and reserve agent reasoning for the moments where judgment actually changes the outcome.

🖥️ OpenRouter gives any model a hosted shell

Need your chosen model to do more than talk? OpenRouter’s new beta Shell tool lets any model on the platform run commands in a hosted Linux container and move files through a Files API. Read the original post on X →

• Add `openrouter:shell` to your tools array and the model decides when it needs a terminal.
• The server tool supports both native OpenAI shell and Anthropic bash tool specifications.
• Models can process uploaded file types by writing scripts, then produce requested file types the same way.
• Compute costs $0.0001 per active second, including file usage, with a 30-second minimum for cold container starts.

What it means: Add terminal and file-processing capability to multi-model agents through one tool interface, then meter computer time separately when you price or limit customer workflows.