Warehouses, Agent Files, and Compute Bills Reshape the Entry Rung

Entry-level knowledge work is moving along a path that warehouses, agent tooling, and AI infrastructure have already taken. In each field, output grows while fewer people start at the bottom.

Key points

• Junior knowledge work is following warehouses toward more output per worker and fewer entry hires.
• In both fields, headcount falls mainly through reduced hiring and only rarely through layoffs.
• Firms are writing junior know-how into Markdown files that instruct agents.
• Hours-long agent runs cost about 10 million tokens each and will likely pay off first in high-value work.
• Fixed fees and outcome pricing push the cost of failed agent work onto firms.
• The shared risk is that firms promise training they never fund and no one audits.

Where the lines cross

Warehouse automation and junior hiring

Amazon's average headcount per facility fell to about 670¹. That is the lowest level in 16 years. Packages shipped per employee rose from 175 in 2015 to about 3,870¹. Industrial robot installations have stayed above 500,000 a year for four straight years².

Knowledge work shows the same pattern in its payroll data. Employment of workers aged 22 to 25 in AI-exposed occupations sits 19% below trend³. Reduced hiring drives the gap. Experienced workers in the same occupations show no comparable decline³. A Dallas Fed paper finds that AI exposure lowers entry-level hiring and raises experienced wages in occupations with a high experience premium⁴.

Both fields shrink at the entry point without mass layoffs. Amazon's planning documents describe fulfillment designs built for robots, with humans handling only exceptions¹. One general counsel now drafts motions to dismiss with AI and no longer needs a first-year for that work⁵. We judge it likely that more legal and finance teams adopt this design, with agents doing the routine work and seniors checking the output.

Both fields also make training promises that are hard to verify. Amazon's upskilling claims resist audit¹. Law firm leaders say first-years are not going away. In a budget exercise at the same forum, nearly every table cut associate pay and kept the AI platform⁵.

Agent context files and junior training

Developers configure coding agents mainly through Markdown context files kept in the repository⁶. AGENTS.md is emerging as the shared standard⁶. Cache reads cost about one-tenth of base input price⁷. That pricing makes stable instruction files the cheapest tokens an agent reads⁷.

Davis argues that entry-level work rests on codified knowledge that AI handles well⁴. A context file is codified knowledge written for a machine. The first-year manual is becoming the agent's instruction set. The Model Context Protocol now sees close to 500 million SDK downloads a month⁸. Agents can now reach the document systems that juniors once worked through by hand.

The same file could teach a human junior. Ropes and Gray already credits 20% of first-year billable hours to learning AI in daily work⁹. The risk is brittleness. Static prose instructions cannot be verified, and agents follow them less reliably as files grow longer⁶. A first-year who learns to write context files may never learn to read a contract⁹.

Compute costs and the price of junior hours

On Long-Horizon-Terminal-Bench, agents used about 9.9 million tokens and 85 minutes per attempt¹⁰. The best model fully passed only 10.9% of tasks¹⁰. METR puts the best agent's 50% time horizon near five hours¹¹. Large tech firms spent over $400 billion on capital expenditure in 2025 and expect a further 75% rise in 2026¹².

Clients are moving prices toward outcomes. Citigroup asks bidding law firms to show AI savings¹³. Zendesk bills $1.50 per automated resolution¹⁴. Outcome pricing arrives while agents still fail most hours-long work. Under a fixed fee, the firm pays for every failed agent run and for the senior time spent checking it. Junior hours used to cover that checking. Fixed fees now leave it unbilled¹³. Our Contradictions piece traced this gap between task horizon and benchmark failure.

Open-weight models trail the closed frontier by about four months¹⁵. Cheaper capability will plausibly reach small firms within a year of reaching large ones. We expect large firms to cut junior intake first and small firms to follow.

What the overlap opens up

Warehouses show that entry headcount can fall for years while output rises. Knowledge work adds a cost warehouses never faced. Junior tasks in law and finance were also how juniors learned the craft. Together, these trends point to a thinner mid-career bench by 2030³. They also show how training could be rebuilt. Firms could turn the agent instruction files they already maintain into teaching material. They could also fund simulation the way airlines fund pilot simulator hours⁵. We put it at about 30% that most large law and finance firms budget training this way by 2030. The likelier outcome is that only a minority funds it and the rest let it lapse.