Harvard: agents write 30% more code, no more software — plus Meta's Muse Wonder, OpenCode 2.0 and the Codex daemon row

ArtCraft — is a set of seven open-source desktop apps written from scratch in Rust with Claude Opus 5.5, copying Photoshop, Illustrator, Premiere and the rest, with Word, Excel and PowerPoint imitations following. It is the demo clients will send you as proof that one person and an agent can now rebuild commercial software over a weekend. The image editor has close to 52,000 GitHub stars, yet the project itself concedes it is not ready for daily professional work, and reviewers found it very buggy. Trending Topics

Pensieve — is a free macOS app that indexes your files, git repositories and past agent sessions on your own machine, then hands the matching passages to a coding agent over MCP, the standard way tools plug into agents. Context waste is a budget line: agents search and read dozens of files before they find the one that matters. In the developer's own test each lookup took two tool calls instead of 3.4, and it runs only on Apple Silicon Macs with macOS 14 or newer. tedhaley.ca

OpenCode 2.0 — is the open-source coding agent rebuilt so a single background service sits behind its terminal, desktop and web apps: begin a session in the terminal, carry on with it in a browser. Worth knowing if a client standardised on it to stay independent of any one model vendor — this upgrade is not a free ride. Plugins written for version 1.0 stop working, and updates are no longer installed for you; the major upgrade has to be run by hand. The New Stack

“The old API grew haphazardly as we and the world collectively learned what agents are and should be. It was bad.”

Dynamic workflows — let Claude design its own team of sub-agents for a job instead of you configuring a coordinator and a fixed roster; they live in Claude Console, the interface for customers using the API. This is the "who manages all these agents" question clients keep asking, answered by handing the management layer to the model itself. In Anthropic's own side-by-side a flow cost much more than a single agent but found far more; the model is not told its budget while running, partly because knowing would make it anxious. Latent Space / The New Stack

Agent37 — is a Y Combinator-backed service renting throwaway machines — sandboxes — for running agents, priced well under the established providers. If you are costing out an always-on agent per customer, this is the new floor to negotiate against; commenters on the launch thread expect a race to the bottom and doubt anyone keeps the margin. Customers average $0.37 a month per sandbox, packed onto rented Hetzner and OVH hosts with gVisor, which rules out Docker-in-Docker and Android emulators. Hacker News

Muse Wonder — are Meta's first models it describes as frontier-grade, built for agents that remember a user across conversations and keep going on long tasks; the larger version is restricted to Muse Code subscribers and no weights have been published. After Llama 4, Meta had dropped out of the serious vendor conversation — clients will ask whether it is back in. Through the API it is billed at $4 per million tokens in and $20 per million out; training ran at Meta's Prometheus site in Ohio. TechCrunch

Decart — Amazon is reported by the Wall Street Journal to be in advanced talks to buy the Israeli startup behind real-time video editing and generated 3-D environments for roughly $7 billion. Neither company has confirmed it. Anthropic looked at the same company at $6 billion in August and walked away in September after diligence — a useful read on how differently buyers value this layer. Decart has raised over $400 million and was valued near $4 billion in May, in a round Nvidia joined. Techstrong.ai

OpenAI fires three safety researchers; they dispute why. OpenAI dismissed Tomek Korbak, Mikita Balesni and Jasmine Wang over what it calls mishandling of sensitive information. The three published a letter saying that reason is a pretext and they were pushed out for safety advocacy; OpenAI says there was a further breach of trust it has not described. Vendor governance is now a client question, and this is the episode people will cite when asking whether a lab's internal critics can speak. Their letter asks OpenAI to embed outside safety auditors and keep frontier models monitorable — requests the company says it strongly agrees with. Don't Worry About the Vase

AI Security Institute — Britain's model-testing body will again see OpenAI and Anthropic models before release, reversing a White House request last month that US officials check them first. For UK and EU clients, state pre-release testing is quietly becoming part of the assurance story procurement teams lean on. Anthropic had withheld its Mythos model from the institute; the minister also launched a task force whose two pilots turned up over 1,000 possible weak spots in government systems. The Next Web

Spotify — merged roughly twice as many code changes a month year over year — about 8,100 up to 17,000 in August — on the back of AI-assisted development, and has been reworking testing, rollout and rollback to cope. The counterweight to the Harvard study: throughput can rise, but only where the verification side gets the same investment. One automated dependency upgrade passed every existing check and still failed in production; automated changes are now scheduled inside the owning team's working hours. InfoQ

QGIS — the widely used open-source mapping project has approved a flat ban on AI-assisted contributions, merged on 7 October, after a gentler human-in-the-loop policy failed to hold in practice. If your client's engineers contribute upstream, contribution policy is turning into a compliance item rather than etiquette. Maintainers counted around 50 pull requests tagged as AI-assisted, roughly 40 of them closed without merging; the author frames the ban as a closed door to be opened later, case by case. GitHub (QGIS)