Three Open Questions on Agents and the Junior Workload

Agents now complete five-hour tasks about half the time and still fail most real professional projects¹². Employers appear to be cutting entry hiring on the first number alone³.

Key points

• Top agents complete five-hour tasks half the time and still fail most multi-hour professional projects¹⁴.
• Employment of young workers in AI-exposed jobs sits 19 percent below trend for reasons still in dispute⁵.
• Buyers expect fixed fees and per-outcome pricing well before sellers report any price pressure⁶⁷.
• Few firms fund the training that billed junior work once paid for⁸.
• We judge it likely that agent capability arrives before firms rebuild the junior training path.

The tensions

Our signal scan found agents reaching the junior workload before firms rebuild training. Three conflicts in the evidence decide how fast that gap closes.

Can agents already carry a junior's daily workload?

The case for is steep. METR's Time Horizon 1.1 puts the best agent at 320 minutes on tasks it completes half the time¹. That horizon has doubled every 131 days since 2023¹. Most discrete first-year assignments in software, law, and finance fit inside five hours. The Remote Labor Index rose from 2.5 to about 16 percent of paid freelance projects in eight months².

The case against sits in the same data. Agents still fail five of six Remote Labor Index projects². Mercor's APEX-Agents benchmark opened at a top score of 24 percent². Inconsistency was its main failure mode². On Long-Horizon-Terminal-Bench the strongest model passed 10.9 percent of multi-hour tasks perfectly⁴. Each attempt consumed about 10 million tokens⁴. METR calls its own intervals very wide¹. It publishes no 80 percent reliability horizon¹. A 50 percent success rate needs a human checker. That checking is the work juniors once learned from.

An 80 percent horizon above two hours would settle the question for discrete tasks. A Remote Labor Index score above half would settle it for whole projects. We think the first is plausible by late 2027 on the current trend.

Is AI the reason firms hire fewer juniors?

The case for is broad. Stanford's Canaries update finds employment of 22 to 25 year olds in AI-exposed occupations 19 percent below trend⁵. The prior version reported 13 to 16 percent⁵. Experienced workers in the same jobs show no comparable gap⁵. A Dallas Fed study of 205 occupations finds that AI exposure lowers entry hiring where experience pays most⁹. The same exposure lifts experienced wages⁹. Among CEOs, 43 percent now plan to move away from junior roles, up from 17 percent a year earlier³.

The case against concerns cause. The Canaries authors call their results descriptive⁵. UK trainee solicitor places fell only 1.7 percent¹⁰. Legal Cheek credits most of that drop to solicitor apprenticeships¹⁰. The paper assigning AI about 30 percent of the falling college wage premium is not peer reviewed¹¹. Tyler Cowen doubts AI is the main driver¹¹. CEOs reporting the best AI returns were more likely than laggards to add juniors³.

Firm-level data linking agent adoption to entry hiring would settle the question. New disclosure rules will not supply it. New York, California, and Connecticut now require WARN filers to state AI's role in layoffs¹². WARN covers mass layoffs only¹². A firm that quietly stops hiring juniors files nothing¹².

Will clients stop paying for junior hours?

The case for comes from buyers. In the Wolters Kluwer survey, 62 percent of legal department respondents expect AI to significantly reduce reliance on the billable hour⁶. Citigroup asks bidding firms to show AI savings⁶. Morgan Stanley plans more fixed fees⁶. Software pricing has moved further. Zendesk bills 1.50 dollars per automated resolution⁷. In a Cruxy survey, 97 percent of SaaS CEOs expect to leave seat pricing within two years⁷.

The case against comes from sellers. In the 8am report, 83 percent of 1,300 legal professionals see no client demand for AI price cuts⁶. Of the same SaaS CEOs, 94 percent say seat pricing still fits their product⁷.

The share of matters billed at fixed fees would settle the question. The stakes are training. A fixed fee removes junior hours first⁶. Partner profit then pays for training⁶. Ropes and Gray credits a fifth of first-year billable hours to AI learning¹³. At a Thomson Reuters forum, most firm leaders protected AI budgets and trimmed associate pay⁸. The real test is whether credited learning hours survive the first down year.

Why these questions matter

Base headcount on measured agent reliability. Trend lines are a weak basis for hiring cuts. Keep a junior cohort sized to the review work that half-reliable agents create. A minority of law firms already budget training like pilot simulator hours⁸. Finance shows the cost of waiting. Senior accountants are now the hardest finance role to fill³. Every firm that buys seniors externally assumes another firm trained them.