Anthropic says juniors fell to 9% of tech hires
Meeting time per employee rose 20%, and managers increasingly valued judgment over implementation skills.
By Saeed Ezzati8 min read
The audio edition
Listen to this newsletter
0:006:00
Read transcript
Anthropic’s junior employees made up 9% of its technical hires, down from 27%. The company says technical hiring still grew in absolute terms; sales, finance and recruiting simply grew faster, bringing technical departments’ share of employment from 61% to 56%. That shift comes with a change in what managers say they want. By April, about a third of machine-learning managers and a fifth of software-engineering managers reported hiring less because of AI. Hiring managers increasingly emphasized autonomy and judgment over specific technical expertise and implementation skills. That’s a striking picture of a workplace where AI can produce more, but people still have to decide what matters and check the result. Anthropic’s study, published October 16, combines employment records, AI-use data, output measures, surveys, calendars and hiring decisions with 17 employee interviews. It describes one highly automated AI company—not an economy-wide hiring trend. The volume of agent work helps explain the management challenge. By August, employee-directed interactive agents logged six times the hours of their handlers, using a 40-hour employee week as the comparison. That is runtime, including tool use, not six times as much equivalent human work. Anthropic estimates that even non-technical employees would need half their workweek just to read everything their Claude agents produce. In an April survey, one in five employees said their job had shifted from doing work to managing Claudes. Output measures are striking, but they need context. Anthropic says code commits per engineer reached five times their pace six months earlier. In a comparison of 192 engineers who adopted coding agents with 74 who didn’t, code additions rose after adoption. But adoption was voluntary, so the researchers say the comparison may be biased. Human- and agent-written code showed no difference in subsequent removal or reversal rates—useful checks, not a complete measure of quality or business impact. And automation hasn’t cleared the calendar. TechCrunch’s account says meeting time rose 20% year over year per employee, and 16% for technical contributors. Anthropic points to trust, accountability and face-to-face coordination. The practical question is not just how much work agents can produce, but how much human attention it takes to direct and verify that work. That same tension between AI-generated work and human coordination shows up in OpenAI’s new shared presentation workspace for ChatGPT. Eligible Pro, Business and Enterprise users with Space access can create a deck from a conversation, uploaded files or an outline, and can provide an example deck to guide its style. Once it’s created, teammates can edit the same presentation in real time and leave shared comments. Revisions can be made directly—changing text, images, shapes or charts—or requested from ChatGPT. Users can also generate custom images or add their own. The aim is to combine AI-assisted creation with hands-on editing and team feedback, rather than making every change another prompt. The feature is rolling out on the web and desktop, but a paid subscription alone doesn’t guarantee access; users also need Space access. Web users can export to PowerPoint, PDF or Google Slides. On desktop, they can open a deck in PowerPoint or Keynote if the relevant app is installed. The useful detail to watch is whether teams can move smoothly between AI suggestions, direct edits and review without losing track of who changed what. The question of how far agents can coordinate without constant human direction gets a more controlled test in Minecraft. Google DeepMind researchers report that four Gemini 4 Argon agents defeated the game’s Ender Dragon in 27 hours and 14 minutes, starting from a random world. The agents had one objective, but no researcher-assigned game plan or human input during gameplay. They communicated through in-game chat and shared notes, divided up jobs, and wrote and ran 2,783 programs. One agent scouted for supplies, another mined for diamonds, while others made tools and prepared a portal. The run included 605 player deaths. The agents controlled characters through Mineflayer, a JavaScript library, by writing programs for actions such as walking, mining, crafting and fighting. The boundaries matter: this was Minecraft Java Edition in survival mode on Easy difficulty, and researchers could pause or stop the run. The authors say an observer agent checked the players’ claims against the record and found no safety incidents. They present the result as a sandboxed demonstration of sustained planning and coordination—not evidence that agents can reliably handle enterprise work. The interesting comparison is less “agents can do anything” than how they preserve plans, share information and recover from failure over time. And in research, Anthropic says Claude Mythos 5 designed circuits that used fewer resources across a set of quantum-computing tasks. An automated evaluator checked the proposed circuits and scored them using qubit count multiplied by gate count. Anthropic reports reductions in both measures on seven tasks; an eighth produced a new design with different tradeoffs. One result stands out: Anthropic says Claude cut the gate count for a decoder used in decoded quantum interferometry by a factor of 12 compared with Google’s earlier construction. The work also tackled a core quantum-circuit tradeoff: storing temporary information uses more qubits, while reconstructing it later can require more gates. Claude proposed designs that navigate that balance in different ways. These are resource-count improvements, not measured runtime speedups. The score is a proxy for computational cost, and the results don’t establish that useful quantum computers are close. Across today’s stories, the shared constraint is clear: more capable systems can generate work, coordinate actions or improve a design, but the gains still depend on how humans evaluate them—and on whether the result holds up beyond the test. Watch for evidence of that validation, not just impressive output or a better resource score.
