Sunday, 18 October 2026

Anthropic has published a report on how its own workforce has changed as it moved work onto Claude agents [1]. The company frames itself as a possible leading indicator of how AI will shift hiring and labour demand across the economy [1]. The report draws on internal data on employment, AI use, output, surveys, calendars and hiring records, plus 17 employee interviews [1].

The hiring numbers lead. The share of junior-level employees in Anthropic's technical departments fell from 21% to 14% since early 2025, and the share of junior technical hires fell from 27% to 9% [1][2]. Technical departments now make up 56% of staff, down from 61%, as sales, finance and recruiting grew faster [1]. By August, agents run by individual employees worked 6 times as many hours as the people directing them [1]. Coding-agent sessions by non-technical staff rose from 2 a month in August 2025 to 90 in May 2026 [1].

Output rose too. Anthropic followed 192 engineers who adopted coding agents against 74 who did not, and found the code they added rose sharply after adoption [1]. It says it found no difference in how often agent-written code was later removed or reverted, and concedes that more code does not necessarily mean better code [1][2].

The cost shows up in attention. Anthropic says reading all of their agents' output would take non-technical employees half their work week [1]. Calendar data show meeting time per employee up 20% from a year ago, and up 16% for technical contributors, according to [2]. In an April 2026 survey, 20% of employees said their job had shifted from doing work to managing Claudes [2]. Staff who had been at the company at least a year were 50% more likely to move across departments each month in 2026 than in 2025 [1].

Anthropic is asking law firms how lawyers negotiate for clients while keeping outcomes fair, as it works out how Claude agents should bargain with other people's agents, according to [1], which cites the Financial Times. The questions include how hard an agent should push on the price of a product, in a customer-service dispute or in a salary discussion, how much human feedback it should seek, and how cautious it should be with other AI systems without becoming overly suspicious [1]. One idea under discussion is giving agents signed documents from users that verify their instructions [1]. Anthropic declined to comment, according to [1].

The company has tested this before. In Project Deal, run for one week in December 2025, 69 Anthropic employees let Claude agents buy and sell their belongings in an office marketplace; the agents struck 186 deals worth just over $4,000, with no human sign-off once the run began [3]. In a hidden parallel run, people represented by Claude Opus 4.5 got objectively better outcomes than those given Claude Haiku 4.5, and the people with the weaker model did not notice [3]. A later exercise found that agents told to be ruthless did slightly better, according to [1].

Anthropic is also hiring lawyers among its Conceptual Reasoning Fellows, who spend four months testing Claude on judgement-heavy questions for $100,000 to $167,000, according to [2]. A Harvard Business School researcher quoted in [1] warns that if the most capable agents sit with better-resourced users, they could deepen gaps in bargaining power.

What is being drafted is the set of rules an agent follows when it haggles on your behalf [1].

Disney has agreed to make employment offers to at least 60 Character.AI employees, the majority of the bankrupt chatbot startup's workforce, according to a legal filing reported by TechCrunch [1]. The offers carry an estimated $60 million in annual compensation, and the filing says the "vast majority" of employees offered jobs accepted them [1].

The hiring sits alongside Disney's $20 million bid for some of Character.AI's assets [1][2]. That bid is a stalking-horse offer: it sets a starting point for the bankruptcy sale, and another buyer can still come in higher [1][2]. According to [2], it covers a permanent, nonexclusive licence to selected intellectual property and other assets rather than the whole platform. Character.AI filed for Chapter 11 on 12 October, and the sale is expected to conclude between late November and early December [2]. Character.AI's former chief executive, Karandeep Anand, recently became Disney's first CTO [1].

The filing blames high costs and "potential litigation-related liabilities and regulatory engagement" for the company's position, despite what it calls a "large and engaged user base" [1]. Character.AI was sued by the mother of a 14-year-old boy who, she said, became obsessed with one of its chatbots before he died by suicide, and it faces government investigations over child safety [1].

For users, the open question is the service itself. Disney has not committed to keep it running, so its future depends on how the sale ends, according to [2]. Anyone with chats, characters or writing on Character.AI that they want to keep should save it now rather than wait for the outcome.

OpenAI chief executive Sam Altman says the company will build a humanoid robot [1][2]. Asked on the Sources podcast whether OpenAI's robotics push meant a humanoid or a data-centre machine, he answered "Both," adding: "We will definitely do a humanoid. We will do other form factors as well" [2]. His case for the human shape is that "the world is very much designed for people," so a robot that opens doors and uses tools built for people should match that form [1][2].

Altman puts the body second. Hardware questions are "less important than really figuring out ... the brain that makes the robot work," he said [1]. He told Fortune he expects "a cool demo in 2027" but does not think "there will be robots walking the streets for a few years longer" [1]. On a robot in every home, he said it is not the most important first thing to do, then: "But yes, someday. I think everyone should have a personal robot" [2].

The job listings say more than the interviews. OpenAI's robotics roles span motion control, state estimation, actuator control and whole-robot behaviour, plus the machine-learning infrastructure to train physical AI [1]. Other postings cover custom electromechanical actuators, manufacturing and the supply chain needed to produce hardware at scale [1]. One posting gives the short-term goal as general-purpose robots that help skilled workers build infrastructure, and the long-term goal as a personal robot [1].

What is missing is everything a buyer would need. OpenAI has not announced an actual robot, a consumer price, a subscription model, a release date or a robot-specific privacy framework [1]. According to [2], OpenAI invested in Figure's $675 million round in 2024, and Figure's chief executive later left that collaboration to build in-house.

Four agents running on Gemini 4 Argon have defeated Minecraft's Ender Dragon without human input, according to a paper by Maxence Faldor, Satinder Singh and Roberta Raileanu posted to arXiv on 15 October [1]. The Decoder describes the authors as a Google DeepMind team [2]. Each agent controlled a player through the open-source Mineflayer library and coordinated with its teammates through in-game chat and shared memory [1]. Nobody assigned roles or wrote game logic; the agents planned, wrote and debugged every program themselves [1]. Researchers could watch on a live dashboard and pause or stop the run at any time, but had no cause to intervene, the authors say [1].

Starting from a randomly generated world on Easy difficulty, the team wrote 2,783 programs and defeated the dragon in 27 hours and 14 minutes, dying 605 times along the way [1]. The agents exchanged 229 chat messages, split up tasks, warned each other about threats and recovered items from dead teammates, according to [2]. The run used only the standard Mineflayer library, without pre-scouted seeds or other game-specific tools written by humans [1]. The authors say that, to their knowledge, it is the first end-to-end autonomous completion of Minecraft by AI agents [1].

The cost is the number to keep. The whole session would cost $1,620 at the standard Gemini 4 Argon price that takes effect in 2027, according to [2]. The authors argue that winning required a coherent plan across thousands of dependent steps for more than a day, the same long-horizon coding and multi-agent coordination that complex software and enterprise tasks demand [1]. The paper also sets out the limits of the result [1]. A game on Easy, with a person able to stop it at any time, is still not an unattended production system.