What Would Have to Be True for Agentic Coding to Replace Junior Engineers

I read every major model release. Most of them ship a coding number.

The number goes up. The conclusion everyone draws is that junior engineers are finished.

I think that conclusion is being reached the wrong way. People are reasoning from a benchmark score to a labor market outcome, skipping every step in between.

So let me do it differently. Instead of asking “will agents replace juniors,” I want to ask what would have to be true for that to happen. Then check each condition against the best evidence available.

There are four. Three of them are not met. The fourth is the one that should worry you, because it does not require the other three.

The best measurement we have here is METR’s time-horizon work. They time human experts on real software tasks, then find the task length at which a model succeeds 50% of the time.

The main result is that this horizon doubled roughly every seven months from 2019 to 2025. METR’s updated Time Horizon 1.1 expanded the task suite by 34% and doubled the count of tasks running eight hours or longer. Independent readings of the 2024 to 2026 window suggest the doubling has since accelerated . The live leaderboard now puts frontier horizons in the hours.

That sounds decisive. Read the methodology and it stops being decisive.

That is precisely the wrong shape. A junior engineer’s first six months are almost entirely context acquisition. Which service owns this. Why that abstraction exists. Who to ask. The benchmark measures the one part of the job that has been stripped of the thing that makes it hard.

In February 2026, OpenAI stopped reporting SWE-bench Verified and recommended others do the same.

Their reasoning is worth reading in full, but two findings stand out. They audited a 27.6% subset of the dataset and found that at least 59.4% of the audited problems had flawed test cases that reject functionally correct solutions. And they found contamination: frontier models could reproduce exact gold patches and verbatim problem details, indicating training exposure.

State of the art had moved from 74.9% to 80.9% over six months. The question OpenAI asked was whether the remaining failures reflected model limits or dataset properties. The answer was mostly dataset properties.

Move to a harder, less contaminated set and scores fall off a cliff. SWE-bench Pro was built for exactly this, and frontier performance on it sits far below the Verified figures the launch posts advertise. Newer suites like Terminal-Bench and long-horizon evolution benchmarks are being built for the same reason.