Four Gemini 4 Argon agents beat Minecraft without any human help

The DecoderOct 18, 2026Oct 18, 2026, 8:39 AM AST

Google DeepMind put four Gemini 4 Argon agents into a randomly generated Minecraft world without giving them jobs, plans or special game tools. The agents divided the work themselves, passed warnings and recovered gear from fallen teammates, writing 2,783 short JavaScript programs across 27 hours to kill the Ender Dragon. The attempt cost $1,620 and the agents died 605 times.

Summary written with AI by HowTechMoves, based on reporting by The Decoder.

All tech news

AI
MIT Technology ReviewOct 19, 2026
AI researchers stranded Claude on a virtual desert island. It started to “play.”
Researchers left copies of Claude Opus 4.7 alone on a simulated island for 30 hours, able to move an avatar by writing instructions. The agents set their own goals, invented games and kept notes in files that acted as long-term memory. Agents that had played later performed better on set tasks than fresh copies, suggesting undirected play can aid learning.
Read original at MIT Technology Review (opens in a new tab)
AI
EngadgetOct 18, 2026
How to use Google's SynthID AI detection website
Google's SynthID hides an imperceptible watermark in images, video and audio made by Gemini, and its new Detector scans a file for that signal. Upload a file to SynthID.com and it returns a yes or no, though it only finds watermarks from Google and partners such as OpenAI and Nvidia, not every AI model. Users get about 10 checks a day to stop people probing how to remove the mark.
Read original at Engadget (opens in a new tab)
AI
Ars TechnicaOct 18, 2026
Anthropic gives outside AI watchdog “employee-like” access
Anthropic has given the nonprofit evaluator METR a team inside its office, with access to its hardware and internal systems, to test Claude models for dangerous capabilities. METR can publish findings without Anthropic's approval. The arrangement is meant to work like bank examiners who sit inside large banks, letting outside experts see risks that companies cannot share publicly.
Read original at Ars Technica (opens in a new tab)
AI
The DecoderOct 18, 2026
Anthropic has nearly stopped hiring junior staff because AI agents now work six times as many hours
Anthropic's internal study says junior workers fell from 27 to 9 percent of technical new hires since early 2025, while its AI agents logged six times as many hours as employees by August 2026. Agents now handle tasks such as writing code and finding bugs. The company says the results may not apply elsewhere, and researchers warn that fewer entry-level jobs could mean fewer experts later.
Read original at The Decoder (opens in a new tab)

Related stories