Agents outwork staff at Anthropic, DeepSeek adds web search, and Claude heads to Canada

Anthropic published a technical report on October 16, titled 'How employment is changing at Anthropic', written by four economists including Peter McCrory and Maxim Massenkoff. It uses the company's internal data: headcount, agent logs, code output, surveys, Slack, calendars and hiring scorecards, plus 17 employee interviews. The headline finding is that in March 2026 the combined hours its AI agents spent working passed the hours its human staff worked for the first time. By August, interactive agent hours for individual employees had reached about six times human hours.

The hiring numbers are just as striking. Anthropic says the junior share of its technical hires fell from 27% to 9%, and juniors' share of technical-department staff fell from 21% to 14% since early 2025. By April 2026, about a third of machine-learning managers and a fifth of software-engineering managers said they were hiring less because of AI. Code commits per engineer reached about five times their level six months earlier. Mentions of candidates' AI tool use in hiring assessments roughly tripled for technical roles and quadrupled for non-technical ones.

The report also describes how the work itself is changing. 'Meta-work', such as building agent tools, writing specifications and monitoring pipelines, rose from 10% to 22% of interactive Claude sessions between fall 2025 and spring 2026. Standing instructions for agents grew from a few hundred in July 2025 to about 8,500 a year later. About 80% of interactive session time is spent with the agent idle, waiting for a human. Reading every word the agents produce would take more than half of a work week, even for non-technical teams. One engineer described 'an army of Claudes, with every one asking for my attention.'

There are important caveats, and the authors list many of them. This is one fast-growing company describing itself, using tools it sells. The 6x figure measures agent runtime, not finished work, and human hours were estimated from headcount at 40 hours a week. Engineers who adopted agents were already on steeper output trajectories, and business growth is a confounder. The junior figures are shares, so they do not show that the number of junior hires fell, and the authors call the junior-employment relationship descriptive, not causal. Survey wording also changed between waves.

Even with those limits, it is the most detailed public look yet at a company that runs largely on agents. The pattern it describes, fewer junior hires, more people managing agents, and more value placed on autonomy and judgment, is likely to spread as other firms adopt the same tools. The authors argue employers should build early-warning systems for at-risk roles and move people internally, but they also say company commitments alone will not fix the slowdown in junior hiring, and that retraining and income support policy are needed too.

Why it mattersIt gives the first detailed numbers on what an agent-heavy workplace looks like, and the skills it rewards (writing instructions, reviewing output, judgment) are ones freelancers and job-seekers can learn now.

Chinese AI lab DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17. It is a snapshot update to the V4.1-Flash model it launched in September, and it is already live on DeepSeek's website, app and API. The company calls it a 'Pareto improvement in performance and cost': stronger on benchmarks, faster at finishing tasks and producing shorter outputs. The weights are published on Hugging Face and ModelScope, and the Hugging Face model card lists an MIT licence covering both the code and the weights, one of the most permissive licences available.

According to the model card, the model has 763 billion total parameters and a 1-million-token context window, and supports tool calling, a thinking mode with adjustable reasoning effort and image inputs. DeepSeek's own benchmark table shows Terminal-Bench 4.0 rising from 31.2 to 42.3 compared with the previous version, DeepSWE v1.1 from 74.2 to 77.0, and Humanity's Last Exam without tools from 36.8 to 42.6. The company says it trained the update with reinforcement learning to get better at building whole projects end to end, showing demos such as a web racing game.

The bigger practical news may be the new Search and Fetch APIs, which let developers give any app live web search and page reading powered by DeepSeek's in-house search engine. Each endpoint costs $0.75 per 1,000 calls. DeepSeek says calls made through server-side tools in its own DeepSeek Harness, Claude Code and Codex are currently free, with billing starting on October 31 at 16:00 UTC. The model is also optimised for Tencent's CodeBuddy and WorkBuddy tools.

Caveats apply. All benchmark scores are reported by DeepSeek and had not been independently reproduced at the time of writing. The model card's own comparisons show Anthropic's Opus 5 ahead on several tests, so this is not a clean sweep over closed frontier models. At 763 billion parameters, self-hosting needs a serious multi-GPU server, not a laptop. Rate limits and how failed search calls are billed are not yet documented, and organisations with data-sovereignty rules will need to think about routing data through a Chinese provider's API.

For builders, the combination matters more than any single score. Open weights under MIT mean the model can be hosted by third-party clouds and fine-tuned without licence worries, while cheap search makes research and monitoring agents far less expensive to run. Each release like this keeps pressure on the prices of US closed models and gives small teams more room to build agent products with thin margins. The free search window before October 31 is a low-risk chance to test whether it fits a real workflow.

Why it mattersA permissively licensed, near-frontier open model plus cheap web search lowers the cost of building research and coding agents, especially for small teams with tight budgets.

Anthropic, the maker of Claude, has signed a long-term lease for data-centre capacity at Point Tupper, near Port Hawkesbury in Nova Scotia, Reuters reported on October 16. It is the company's first data-centre agreement in Canada. The developer, EverWind, first described its tenant only as a 'leading frontier AI developer' before Anthropic confirmed its involvement. Reuters describes the lease as covering the campus's first phase, and reports put its term at around 20 years.

According to Global News, the first phase will provide 95 megawatts of capacity starting in late 2027. EverWind plans a 325 MW expansion targeted for 2028, which would bring the campus to 420 MW, but that stage depends on commercial, technical and regulatory conditions. The total plan is about C$10 billion of private investment: roughly C$8 billion for the data centre and C$2 billion for onshore wind projects. EverWind says the campus will include several hundred megawatt-hours of battery storage, about 2,000 construction jobs and 215 permanent operating roles.

EverWind CEO Trent Vichie called it the largest private capital investment in Nova Scotia's history. Jacob Glick, Anthropic's head of Canada public policy, said Canadians are using Claude in growing numbers and that the company would meet Canada's bar for responsible data-centre development. The deal comes as AI companies race to secure power and buildings years ahead of need, and as Canada pushes for more domestic, secure computing capacity. The federal government also announced C$100 million the day before for the Wind West offshore wind project.

The project is not without controversy. Cooling is mainly air-based with a sealed water-glycol loop, but water use is still estimated at about 100,000 litres a day in the first phase and about 500,000 litres a day at full build-out; these are design estimates. EverWind says it will fund electricity and water infrastructure so costs are not passed on to local ratepayers. Premier Tim Houston says existing approval processes will apply, while opposition leader Claudia Chender has called for dedicated regulation. Other Canadian provinces and municipalities have pushed back on data centres over power, water and environmental concerns.

For the wider AI market, the deal is a reminder that the next two years of AI capability are being decided now, in leases and power contracts. Capacity in Canada could also help Anthropic serve customers who want data processed inside the country, though the companies have not said how the site will be used. For communities everywhere being pitched large data centres, including in East Africa, the Nova Scotia debate over water, power costs, tax revenue and local ownership is a useful template; EverWind is negotiating a possible ownership stake with a First Nations consortium led by Membertou Development Corporation.

Why it mattersAI capacity is being locked in years ahead through long leases and power deals, and the local fights over water, power and tax revenue are becoming part of the AI story everywhere.