Signal #89: Anthropic cut Junior hires from 27% to 9%

Tuesday edition, friends.

Every senior engineer in this region was once a junior somebody took a bet on. This week the numbers say fewer companies are placing that bet, starting with the lab that builds the agents.

The data comes from San Francisco, Zagreb and Prishtina, and it points the same way. Then there is the price of the work juniors used to do, and a government plan that trains almost everyone except the people entering the job market. Four stories, five minutes.

On Friday, Anthropic published its own employment data. Junior staff fell from 21% to 14% of its technical departments since early 2025, and juniors went from 27% of technical hires to 9%. By August, AI agent hours outnumbered employee hours six to one. Commits per engineer run at five times the rate of six months ago.

The company still tripled in size over the year. The growth went to sales, finance and recruiting, while technical departments fell from 61% to 56% of staff. The agents did not take the senior jobs. They took the work seniors used to hand down. That work is how a junior became a senior: the small bug, the boring test, the first review. When an agent does it in minutes, nobody budgets for the slow human version.

If you run an engineering team, count how many people joined it as juniors in the last 18 months. If the answer is zero, you are spending down a pipeline you are not refilling. Pick one workstream where a junior pairs with an agent instead of being replaced by one, and measure what they learn, not only what ships.

Home furniture brand Blu Dot blew up on CTV with help from Roku Ads Manager. Here’s how:

After a test campaign reached 211,000 households and achieved 1,010% ROAS, the brand went all in to promote its annual sales event. It removed age and income constraints to expand reach and shifted budget to custom audiences and retargeting, where intent was strongest.

The results speak for themselves. As Blu Dot increased their investment by 10x, ROAS jumped to 2,308% and more page-view conversions surpassed 50,000.

“For CTV campaigns, Roku has been a top performer,” said Claire Folkestad, Paid Media Strategist, Blu Dot. “Comping to our other platforms, we have seen really strong ROAS… and highly efficient CPMs, lower than any other CTV partner we've worked with.”

Using Roku Ads Manager, the campaign moved from a pilot to a permanent performance engine for the brand.

Learn More

Ericsson Nikola Tesla has about 2,000 engineers in software development, and CTO Jelena Jurišić told Netokracija this week that the company runs 46 AI agents across the whole lifecycle: requirements, design, development, testing, delivery and maintenance. Teams are getting smaller while shipping more features, and she expects frontend and backend roles to merge. Her rule: you can hand an agent the work, not the responsibility.

The new Kosovo IT Barometer from STIKK (full disclosure: I sit on its board) shows the same shift further down the market. All 102 firms surveyed used AI tools in 2025. 46% said AI had already reduced the need for some roles, and 87% of those named intern and junior positions. Yet 90% still find it hard to hire skilled staff. The region is cutting the entry level while short of the people the entry level used to produce.

This hits our outsourcing model directly. Three in four of the surveyed Kosovo firms export, and much of that work is billed as junior hours inside senior-led teams. If you run a dev shop, price one contract by outcome this quarter, and keep a paid junior track inside it so you still have seniors in 2029.

On Saturday, DeepSeek released V4.1-Flash-1017 with open weights. By its own figures it scores 77% on the DeepSWE coding benchmark at just under $0.20 per task, while frontier models from OpenAI, Google and Anthropic land between $2 and more than $10 per task on the same chart. Two days earlier, Reflection put the weights of its 500B-parameter Beam model on Hugging Face, fine-tuning recipes included.

That is the number a manager now holds up against a junior's day. When the first draft of most engineering work costs cents and runs on your own servers, it stops being a training budget. It also kills the old argument that juniors were the cheap option. They were never only cheap. They were the apprenticeship.

Run the test on your own team. Take ten tickets from last month that went to your most junior people and put them through an open model. Whatever it clears is work you should stop paying humans to do. Whatever it fails on is the curriculum your juniors should be learning now.

On Friday, minister Damir Habijan presented Croatia's National AI Plan, worth €932.1M through 2032. The first action plan runs to 2028 with €382.1M across 33 measures, including €139M for compute and data, a Croatian language model and an AI assistant in the e-Građani portal. By 2028 it aims to train 7,000 citizens, 10,000 teachers and 5,000 public servants.

Next door, the six Western Balkan economies got a shared AI roadmap to 2030 at the Herceg Novi digital summit, still waiting for leaders' endorsement and with no funding attached. Croatia brought a budget. The rest of the region brought a framework. Neither names, in its published targets, the group this issue is about: graduates trying to get their first year of real work.

The Croatian plan now goes to a 30-day public consultation. If you run a tech company there, send one concrete proposal: co-funded junior placements where the state covers part of the first year and the company pairs each hire with its agent stack. Everywhere else in the region, take the same idea to your ICT association before the roadmap gets its Berlin Process endorsement.

Dev: Mercury 3 writes at 1,068 tokens a second. Inception's new diffusion model has a 1M-token context, speaks the OpenAI API format and costs $0.20 in and $0.70 out per million tokens, half that at launch, with 100M free tokens to start. Use it as the fast, cheap backend for agent loops where waiting hurts more than raw quality. (Inception)

Productivity: Cloudflare Clairic pulls typed values out of messy text. The preview model on Workers AI returns numbers, dates and text spans, each with a confidence score. In Cloudflare's own claims pipeline it ran about 3.2x faster and cost 49% less than sending everything to an LLM. Point @cf/cloudflare/clairic at your invoice or order inbox first. (Cloudflare)

Marketing: Mistral Timestral forecasts on a plain CPU. Four zero-shot forecasting models, from 5M to 209M parameters, need no training data and returned a day-ahead forecast in 10 to 140 ms on a standard server CPU in Mistral's test. Weights are on Hugging Face under a modified MIT licence. Feed it last quarter's traffic or signups and plan your next campaign around the curve. (MarkTechPost)

Fintech: Belgrade's Curvy is in a global privacy funding round. Curvy Protocol builds private payments that can still prove compliance, including payments made by AI agents. It is one of 13 teams in Octant's Epoch 13 round, with a 100 ETH matching pool split by quadratic funding, so many small votes count for more than one big one. Voting closes 22 October at 02:00 Belgrade time. (Startit)

Next edition soon,
Çelik