What shipped in week 43

The open-source, provider-agnostic AI toolkit hits 1.0 with stable core APIs, 23 provider adapters, support for coding agents like Claude Code, Codex and OpenCode in sandboxes, plus a new experimental agent harness.

Replit Agent can now suggest new or improved workspace skills from your chats and projects, which you review and approve before they land. The same update lets eligible top-up customers keep published apps running when credits run out, up to an overage limit, and adds Stripe Atlas on Replit.

Adds /presets (ten saved prompt and slash-command slots), inline arguments for /model and /permissions, image previews on Kitty/Sixel terminals, enterprise MCP sign-in, and credential masking in remote Bash, plus sandbox hardening.

Schedule a message into an existing agent session or have it start a new one, using a cron expression for recurring jobs or a timestamp for one-offs. Add the automations tool and agents can create, update or remove their own schedules from natural-language requests.

Moonshot's new flagship model is live on the Kimi API at $2/$10 per 1M input/output tokens, plus a 2-3x faster Highspeed tier at twice the price. Moonshot says full open weights follow by 4 Nov 2026; Artificial Analysis scored it 55, behind only Claude Opus 5.5.

The new consolidated GitHub app in Teams brings notifications and Copilot together: @mention GitHub on an issue notification and Copilot investigates, fixes the bug and opens a pull request. The legacy app retires on 16 November.

v3.11.212 rebuilds the Agent Window on the Devin web app UI: up to 8 sessions side by side, per-session Files, Changes and terminal views, cloud app previews with an element picker, and Enter to steer a running turn. v3.11.216 (24 Oct) fixes Local sessions and CLI install.

Cursor's agents can now click and type across any desktop app on your computer, not just the browser, for example driving the iOS Simulator to test a feature they just built. The agent uses it when a task needs it, or you can @-mention Computer Use; update Cursor to get it. Desktop-app release notes weren't out yet at announcement.

An Apache-2.0 research runtime from AWS and UC Berkeley's Hydro project that suspends an agent during each tool call and resumes it when the result arrives, so idle agents cost little more than their history and can run on serverless platforms. AWS says 75,000 concurrent agents fit in a Raspberry Pi 5's 8 GB.

Describe an app and Wodby Agent picks a stack, creates a private repo in your own GitHub/GitLab/Bitbucket, builds it in a live workspace with previews, opens a PR and deploys to infrastructure you control or Wodby Cloud.

Run vercel deploy --temporary to get a live tmp.now URL for a project or a single file without signing in. Temporary deployments expire after one hour unless you claim the project via the link in the CLI output.

An autonomous on-call agent for apps published on Replit: it reads production logs and metrics, investigates alerts, validates a fix and can auto-publish it, then emails you a summary.

A natural-language coding agent in the Microsoft Copilot app that builds apps, dashboards, tools and scripts from a description, using your work context, with a live preview, local or cloud runs, GitHub draft PRs and publishing to Copilot Managed Runtime. Rolling out via the Frontier program.

Google open-sourced its internal design-first AI coding plugin: a long-context 'Elephant' session co-designs a spec, then stateless 'Goldfish' sessions review it before any code is written. Apache 2.0, built for Antigravity first.

Free, Apache 2.0 coding model small enough to run on a laptop (40B-parameter MoE, 2.5B active), so you can code offline and keep your code private. Download on Hugging Face.

OpenAI's frontier models and Codex workflows are coming to Google Cloud's Model Garden, billed via Google Cloud Marketplace; private preview now, GA planned in the coming weeks.

Free VS Code extension with an object database where your code is the schema, so you or your coding agent change a class and commit with no migration scripts. Python only at launch; includes an MCP server.

Anthropic's proactive agent app: runs tasks on its own cloud computer, sends a morning brief from your connected apps, can prototype and build web apps, and asks before sending, paying or deleting. Waitlist for US Pro and Max.

WebSocket support in Vercel Functions is generally available, for AI chat, streaming and realtime multiplayer apps, at normal Function pricing (CPU billed only while processing messages).

New speech-to-text and text-to-speech models (50+ languages) plus a voice agent SDK that listens, reasons, calls tools and speaks in one real-time pipeline. Transcription from $0.0035/min.

Ask Lovable to create a Supabase project for an app with no backend: pick the organisation and region and Lovable creates and connects it, billed to your own Supabase org.

Cursor's Agent SDK is generally available: define an agent as a folder of code (instructions, typed tools, skills, MCP connections, channels), test it with evals, and deploy it to Cursor with agent-sdk deploy or run it on your own infrastructure. Agents can wake on GitHub, GitLab, Bitbucket, Slack, webhook and cron events.

Connect Slack to v0 so it can use team discussions as context for a change and post updates to Slack when you ask. Available on all v0 plans.