📉 Anthropic's AI agents now log 6x more hours than staff

PLUS: 🎨 The best free AI training to level up your game

Eduardo Andreu

5 min read

In this newsletter

This week’s developments

Graph from Anthropic’s report, showing agent hours overtaking human hours among their employed staff

Source: Anthropic

📉 Anthropic's AI agents now log 6x more hours than staff

In summary: Anthropic just published data on its own workforce showing that by August, AI agents logged six times the hours of the staff directing them, while the junior share of its technical hires fell from 27% to 9% since spring 2025.

The details: 

  • Anthropic's chart shows changes and updates to software by engineer up 5.1x since December 2025, with contracts processed per lawyer and queries per data scientist both up 3.1x

  • Meta-work such as agent specs and monitoring rose from 10% to 22% of Claude sessions, and reading all agent output would fill over half a non-technical week

  • The median employee spends 20% more time in meetings than a year ago and technical ICs 16% more, with staff calling face-to-face time indispensable

  • Hiring scorecards that note creative use of AI tools have tripled for technical roles since 2024, with autonomy and judgement the fastest-rising traits

Why it matters: I read Anthropic’s article as a window into the future of employment in AEC. There are two takeaways which AEC firms will mirror First, the value of technical tasks will diminish. Doing an audit of your own week is the easiest way to pinpoint which things you do will loose value. Second, the value of judgement and taking responsibility will increase. So will the value of soft skills, human interaction, and being the person having the conversations no AI agent can hold for you.

AEC's likely new hire: Individuals that combine AI fluency with real-world judgment and responsibility

If you are interested to read similar stories, see also: 

A TKE elevator technician talking to phone, presumably to an AI agent

Source: Microsoft / TKE

🛗 TK Elevator's AI agents brief technicians before arrival

In summary: TK Elevator just detailed an AI app, built on Microsoft Foundry, that tells its service technicians which lift, floor, fault and tools to expect before they enter a building, then debriefs them by voice once the job is done.

The details: 

  • The app sits on a decade of connected-lift data, with nine digital operations centres using machine learning to give every connected unit a health score

  • After each visit a voice agent asks follow-up questions and logs the answers, so the next technician can search what the last one actually found

  • TKE says its 25,000 technicians maintain 1.4M lifts and escalators, which carry 1.5B passengers a day, across more than 100 countries

  • TKE says a 2025 US pilot with connected units avoided 20,000 unplanned visits and cut callbacks by over 40%, though it doesn't credit the agents alone

Why it matters: AEC firms often lose staff midway through multi-year projects, and replacements spend weeks catching up. Making human decisions visible to agents and writing standing material that agents work from, as TKE does or some other way, can ensure project knowledge stays within the firm, and reduce project onboarding time. For example, it could brief a PM before their first meeting or hand an architect a live job with its history attached. The hard part is getting teams to log as they go!

AEC's likely new hire: Design manager, responsible for keeping a project's information well structured and machine readable

Source: Two wireframe drawings of an electric vehicle chassis on a black background.

🔋 GM tested 300,000 battery designs before building one

In summary: GM just pitched a new EV battery as proof that AI and virtual testing can halve car development, after simulating 300,000+ designs over 300 million CPU hours (the equivalent of one computer running for 34,000 years) before building a cell and getting the pack approved in a month.

The details: 

  • GM aims to cut vehicle development from four or five years to about two using AI and virtual testing, though the new cells won't hit volume until 2028

  • The cells use lithium manganese-rich (LMR) chemistry, which GM claims stores 33% more energy than China-dominated LFP cells at a comparable cost

  • GM says Computational Fluid Dynamics simulations runs that once took up to 60 hours now take 60 seconds, and that decades of GM engineering data now feed its AI tools.

Why it matters: Simulations which are common in the design of buildings, and which sometimes required specialist consultants (e.g. rights of light) will be no more than a click of a button in future AEC software, rerun on every option as the scheme develops. But GM's shortcut runs on decades of validated data, and few AEC firms keep theirs that clean.

A 3D render of a teal, low-poly landscape seen from above

Source: MIT

🏝️ MIT stranded an AI agent on an island, and it played

In summary: MIT researchers just published a study in which 13 Claude agents, given a virtual island and no task for 30 hours, stacked towers and invented bowling and basketball, then beat fresh agents at most building challenges they never trained for.

The details: 

  • Each agent, called Eko, ran on Claude Code with Opus 4.7, a text-only view of the island, six actions and memory files that survived hourly resets

  • Asked to build as many five-block towers as possible in an hour, 62% of played agents built more than two, against 15% of fresh ones

  • Six agents found a mid-jump throw the researchers didn't know their simulator allowed, a quirk that was written into the physics code inadvertently

  • Eko's starting file told it that it was curious and got bored easily, though five runs without that file took up a similar range of activities.

Why it matters: Play is an integral part to creativity and design. So integral in fact, that the lack of play of AI generated designs and writing give them away instantly. As innocent as the games Eko created may sound, if AI models could indeed play, they have the potential to become infinitely more creative.

Screenshot of Claude Academy homepage, with a ‘Welcome to Claude Academy’ heading

Source: Anthropic

🎨 The best free AI training to level up your game

In summary: The best resources to learn the basics of AI are free, unless you are after something very specific. The companies building AI models are all giving away better education than most paid resources. And the best part is that these are courses for everyone: whether you are an Architect, a Structural Engineer, a Project Manager, or else.

Here’s a useful reference list of learning resources without paywalls: 

In addition Hugging Face and NVIDIA offer specialist courses for people who want to learn to use open models. These can be a little more advanced:

And for those who appreciate traditional education, Harvard offers a variety of free certifications. These two courses may appear abstract to most people in the AEC Industry, and will make many professionals roll up their eyes. I, on the other hand, learnt a huge amount taking the original CS50:

Trending AI Tools this week

That’s it for today!

See you next week with more exciting updates,

Eduardo - The human behind AEC field notes

London, UK

Read past issues