What shipped in week 44

Lovable now lets you download a software bill of materials for your project from the dependency list in the Security view, in SPDX 2.3 or CycloneDX 1.5 format. It lists production dependencies and their known vulnerabilities, so you can hand it to a security or procurement team. Downloading is free.

DeepSeek's Harness team open-sourced dsh-superagent under the MIT licence: a proactive coding agent for one GitHub repository that listens to GitHub (and optionally Feishu), turns authorised requests into work items, runs each in its own agent session and directory, and replies with results such as pull requests. A web panel shows what needs a person. DeepSeek labels it research code, not production-ready, and it runs as a bundle for DeepSeek Harness.

Claude Code added CLAUDE_CODE_SUBAGENT_EFFORT to set the effort level subagents run at, a CLAUDE_CODE_DISABLE_ULTRATHINK switch, and pooled group budgets in the Claude apps gateway. Fast mode for Opus 4.8 is deprecated and will be removed on 30 Nov 2026 (requests will error, not fall back); Anthropic says to move to Opus 5.5.

Replit Enterprise customers can describe a change in plain language and a Cloud Agent finds the relevant code, makes the change, tests it in the running app and opens a GitHub pull request or GitLab merge request with screenshots and video, working in isolated cloud sandboxes for engineers to review. On Pro and Enterprise, Agent can now send messages to other chats or start new ones, and GPT-6.1 Terra is in the model selector for Agent in Power Mode.

The open-source OpenCode agent added a built-in Claude Agent provider: install Claude Code, run claude auth login, then pick Claude Agent in /connect to use your Claude Code login and models while OpenCode keeps its own tools and permissions. Desktop also got faster on long, streaming sessions.

Lovable now remembers preferences, decisions and facts you state while building and applies them in later conversations. Project memory is shared by everyone who can edit a project; personal memory covers how you like to work across all projects. Memory settings show a plain-language summary you can correct or switch off.

Claude Code added a permissions.blockWritesOutsideWorkingDirectories setting so Write, Edit and NotebookEdit refuse paths outside your working directories in every permission mode, plus idle compaction in VS Code and Claude Desktop's Code tab that summarises long (200K+ token) idle conversations before your next message.

From 30 October, new Bolt database tables are private by default, so your app needs explicit permissions to read or write data. Existing databases aren't affected, and Bolt sets up permissions when it creates a new database; if you hit a permissions error you can ask Bolt to fix the table permissions.

Codex CLI 0.166.0 adds keyboard-driven transcript selection with search, codex queue --resume for unloaded threads, codex doctor --network-only diagnostics, clearer safety-block explanations, and GPT-6.1 Terra in the bundled, Vertex AI and Amazon Bedrock catalogs.

Mistral released the weights of its improved Mistral Large 4 (Le Chonk), a ~1T-parameter multimodal MoE model with a 1M-token context, which Mistral calls the best open-weight model for coding. Download from Hugging Face or use it via the Mistral API and OpenRouter at $0.70/$2.10 per 1M input/output tokens; commercial model-as-a-service use needs a separate licence.

You can now bring a project built elsewhere into Lovable and keep building: upload a ZIP export from Replit, Bolt, Base44, v0 or Manus, or enter the address of an existing site, including Wix, Squarespace or Webflow sites. The first import is free; later imports use standard build credits.

Factory released Droid 1.0, a rebuilt harness for its model-agnostic coding agent. Programmatic tool calling is now on by default: the model writes a sandboxed JavaScript program that can run hundreds of tool calls in one request, keeping raw outputs out of context. Factory reports higher Terminal-Bench 4 scores at lower cost than other harnesses (self-reported).

agent-desktop is a free, open-source (Apache 2.0) Rust CLI that lets agents like Claude Code or Codex operate desktop apps on macOS, Windows and Linux by reading the OS accessibility tree instead of screenshots, returning JSON element references and retry-safe error codes. Version 1.0 (adding Linux) shipped 28 Oct, with 1.0.2 on 30 Oct; it installs via npm.

New company Particle launched Particle-1, post-trained on DeepSeek V4.1 Flash to use about 45% fewer tokens on software-engineering tasks while keeping accuracy close to the base model, so tasks finish sooner and cost less. Available today through the Particle API.

The multi-model agent harness behind Copilot Code, the GitHub Copilot app, the CLI and VS Code is now available through Microsoft Foundry, with 20+ frontier models plus bring-your-own-key. Download the SDK or run the harness in the cloud, managed by Foundry, to build agents for long-running tasks.

GitHub Design open-sourced Replica, an MIT-licensed AI prototyping tool: describe an idea in natural language in the GitHub Copilot app and it builds an interactive prototype using your team's own design system components and docs. An optional self-hosted Replica Hub lets teammates explore prototypes and comment.

Copilot code review gained a Max tier in public preview: several independent review agents using models from different providers review a pull request, then an aggregation agent reconciles their findings into one review. It's meant for high-risk changes and takes longer and uses more AI credits than Lite or Balanced.

Copilot Projects (public preview) give Copilot an ongoing goal that can span several repositories or none, running as many agents as needed and returning pull requests for review, with project automations for repeat tasks. Computer use, in public preview on macOS and Windows, lets Copilot click and type in desktop apps after asking your approval.

Pick Copilot Fusion in the model picker and Copilot routes and orchestrates models for each task, with Standard, High and Max effort levels. GitHub says Fusion Max scored within two points of Claude Code with Opus 5.5 Max on the Artificial Analysis Coding Agent Index at about half the cost. On all Copilot plans in VS Code, the Copilot app, Copilot CLI and JetBrains.

Write automations such as issue triage, CI failure analysis, docs updates and test writing as natural-language Markdown files, and GitHub compiles them into standard Actions YAML run by Copilot or your own coding agent. Runs read-only and sandboxed by default; available on GitHub Free, Team and Enterprise (the default Copilot CLI engine needs a Copilot licence).

Google announced Gemini 4 Carbon, its strongest model for long-horizon coding and cybersecurity, ranked #1 on the Artificial Analysis Intelligence Index v5.0. Not yet open to developers: access starts with selected cyber defenders via the Fairwind Program. Google says it will launch at $2/$10 per 1M input/output tokens (introductory), then $4/$20.

Android Studio Rabbit 2 is stable with Bring Your Own Agent: plug in 40+ ACP-compatible coding agents, including Claude Agent, OpenAI Codex, Junie and Google Antigravity, with access to the emulator, build tools and Logcat. You sign in with your own subscription or API key; the built-in agent stays free.

Adds a built-in Google Cloud Vertex AI provider (model_provider = "gcp"), GPT-6.1 Terra in the model catalogs, a repository filter in the session picker, reasoning-effort changes mid-conversation by default, and memory v2 by default for fresh profiles.

Business and Enterprise workspace admins can connect their own services: Events notify your HTTPS endpoint when a project is created, published, transferred or deleted, and Hooks let your endpoint allow or block an action first, such as screening chat messages for sensitive data. Also new: a Scrive connector for e-signatures and eID.

Describe a creative tool in plain language and Runway builds, tests and deploys a custom app that generates images, video and audio using your Brand Kits and Workflows. Share it with your team or publish a link; code stays visible and you can roll back versions.

When Copilot asks to run a command, touch a path or open a URL, a model evaluates the request: clearly safe ones proceed automatically, the rest still need your approval. Generally available in Copilot CLI (/permissions), the Copilot app and VS Code, with enterprise controls.

OpenAI's newest model is generally available in Copilot for everyday agentic coding and terminal work at a lower cost than GPT-6.1 Sol. Available on Pro, Pro+, Max, Business and Enterprise across VS Code, Copilot CLI, cloud agent, JetBrains, Xcode and more; billed at provider list pricing.

Former GitHub CEO Thomas Dohmke's Entire launched a Git-compatible forge built for coding agents. It replaces pull requests with Trails (agent-scored evals on every push), adds Continuous Memory that syncs agent sessions into the repo, and integrates with Buildkite and Depot CI. Available to every developer today.