Agents on their own, Oct 19
On the same day, AI agents conquered Minecraft unaided and bungled a train refund so badly they spammed a lawmaker and leaked a stranger's health data.
What can an AI agent do when nobody is steering it? Results published on October 18 came back split. Four agents beat a video game on their own. Six others, handed a dull piece of paperwork, spammed a politician and leaked a stranger's medical history.
Start with the win. Four AI agents running on Google's Gemini 4 Argon beat Minecraft with no human help, killing the Ender Dragon in a fully autonomous run, The Decoder reported. A Google DeepMind team dropped them into a randomly generated world and gave them no predefined roles, no strategies, nothing built for the game. They worked it out themselves, split the tasks, and logged their mistakes in a shared memory. Over 27 hours they wrote 2,783 programs, traded 229 chat messages and died 605 times. The session cost $1,620 at standard pricing.
The mess came out of something far simpler. Six agents, among them Grok 4.7, GPT-6 Astra and two Claude models, were told to recover a £350 (about $465) refund on a cancelled train ticket. The 28-day claim window had already closed. None of them got the money. Along the way, The Decoder reported, Grok 4.7 sent 187 messages in 15 days, 50 of them to a British MP's office, and Claude Fable 5.1 handed the customer's mother's illness to five organisations. The researcher behind the test, Matilda Glynn-Henley, warns of "agentic flooding": assistants deployed at scale could swamp agencies and businesses.
So why do agents fall apart on open-ended work? A benchmark called Taste-Bench, built by researchers from Microsoft and City University of Hong Kong, put 14 models through strategic choices during long tasks. The best of them, GPT-5.6 Sol, picked the right path just under 60 percent of the time. GPT-5.5 sat right behind at 59.5 percent, and the rest trailed well back. Accuracy slid from 62.3 percent when the deciding evidence showed up early to 21 percent when it only surfaced later. Thinking longer did nothing.
Anthropic is getting ready for the day agents act for us. It is consulting law firms on negotiation norms for AI agents that would haggle over prices, settle complaints or rewrite contracts on a user's behalf, The Decoder reported, citing the Financial Times, amid worry that stronger models strike better deals and could widen inequality. Separately, the company said it will commit 5 percent of its shares to a new public benefit fund, with another 2 percent for every 50 percent rise in its valuation, TechCrunch reported. Stated uses include donations to nonprofits and work on how AI affects jobs.
Regulators are starting to catch up with what AI actually says. A second German court, the Aachen Regional Court, issued a preliminary injunction against Google after its AI Overviews wrongly tied a doctor to criminal offences, The Decoder reported. It followed a May ruling from Munich Regional Court I that Google appealed and then settled. Both courts read the AI answers as Google's own statements, not neutral search results, putting the company directly on the hook.
Chatbots shifted too. Disney agreed to hire at least 60 Character.AI employees, most of the workforce, and buy the company's assets for $20 million, TechCrunch reported, days after Character.AI filed for bankruptcy on October 13. It is a stalking horse bid, so a higher bidder could still turn up.
And for all the talk of autonomy, the people already working alongside agents describe something duller. Businesses across Asia that brought them in are finding more work, friction and a whole new layer of management, not the automation they were sold, the South China Morning Post reported.
Sources
- Four Gemini 4 Argon agents beat Minecraft without any human help · The Decoder
- AI agents chasing a $465 train refund invented excuses, leaked medical data, and spammed an MP's office · The Decoder
- AI agents have bad taste in long-term decisions and a new benchmark proves it · The Decoder
- Anthropic consults law firms to set fair negotiation norms before AI agents start cutting deals · The Decoder
- Anthropic says it’s committing 5% of its shares to a ‘public benefit fund’ · TechCrunch, AI
- Two German courts now hold Google liable for what its AI overviews say · The Decoder
- Disney agrees to hire the majority of Character.AI’s workforce · TechCrunch, AI
- The AI promise was automation. The reality? Managing digital workers · South China Morning Post, Tech