AI engineering news, this week

Gemini 4 Argon

Google released it to developers and consumers. GitHub says it is now generally available in Copilot (Pro+, Max, Business, Enterprise), with billing at provider list pricing.

Meta Muse Wonder and Wonder Full

New models aimed at coding, long-running agents and computer use. Launch API price: $4 input / $20 output per million tokens.

Mistral Large 4

Public preview of an open-weight model via API. Weights are due at the end of October.

JetBrains Mellum2.1

A fast open model built for coding agents.

More code, not more software

Harvard study of 700+ firms: coding agents raised lines of code about 30% and pull requests about 23%, but review time rose 49% and issue resolution did not change significantly.

Survey: more debugging, a comprehension gap

A survey on complex codebases links AI-generated code to more debugging, higher failure rates and a comprehension gap.

TesterArmy e2e

Open source TypeScript framework. Playwright-style locators and assertions plus agent steps like agent.act(). Same tests for web, iOS and Android.

Kane CLI by TestMu AI

Terminal agent that lets coding agents check web and mobile software in natural language, with hash-verified evidence packs.

Smaller outlet, summary of a launch announcement. Check the vendor before relying on it.

Leapwork Play

AI-native continuous validation layer for Playwright suites, with Claude, ChatGPT, Gemini and Copilot integrations.

Smaller outlet, summary of a launch announcement. Check the vendor before relying on it.

Klarent raises $8M

Zurich startup: LLMs write Playwright scripts, humans review, tests then run deterministically.

Smaller outlet, summary of a launch announcement. Check the vendor before relying on it.

Playwright MCP writing your tests?

A two-week hands-on account of letting an agent click through an app and generate Playwright tests.