Coding agents write twice as fast as humans can review. The bottleneck moved, not away

The 2x mandate worked, then broke

In a preprint posted this week, researchers report on a longitudinal study of an enterprise mandate to double developer throughput with AI tools. The mandate worked, after a fashion: AI adoption led to an immediate increase of about 12% in pull request throughput per developer. But the gains were driven by larger, more complex pull requests rather than increased working hours — and code review load doubled following AI adoption, with automated review coverage eventually exceeding human review. The study's blunt finding is that review has become the primary bottleneck. The agents got faster. The humans did not. m:…0994

A second preprint, introducing the ExpertCrowd-SWE benchmark of 100 expert-authored engineering tasks in unfamiliar codebases, explains why review cannot simply be skipped. Claude Opus 5.5 and GPT-6 Astra lead the benchmark at 67.3% and 63.7% respectively, but no model achieved full credit on all five runs for 80 out of 100 tasks. Agent performance also shows little correlation with the number of steps taken, which ranged from a median of 33 to 166. More effort from the agent does not reliably mean a better result, which is exactly why someone has to check the work. m:…9448

Cheaper repair starts before the model runs

In a preprint, a team proposes FlowPlan, a method that cuts the token cost of coding agents by shifting repair from raw source-code manipulation to abstraction-guided planning using control-flow and data-flow graphs. It reduces token consumption by 26-68% across six benchmark-model settings while maintaining resolution rates comparable to vanilla coding agents. The savings come from replacing expensive post-edit re-reading with explicit structural verification on program abstractions. In other words, the agent stops re-reading the whole file after every change and checks its plan against the structure instead. m:…0799

In a preprint, another study attacks the same cost from the other end: the bug report. Evaluating repair agents on the SWE-Bench Verified benchmark, the researchers found that suggested fixes — in code or natural language — show the strongest positive associations with agentic repair pass probability. Agents benefit less from information traditionally useful to humans, such as reproduction steps, and more from sentences that expose a repair direction. The implication is awkward for teams that spent years writing good bug reports for people: the reports agents need are different documents, and the ones you have are optimized for the wrong reader. m:…0797

What to do about it

Put the two findings together and the shape of the fix is clear. If review load doubles while throughput rises 12%, the answer is not more reviewers — it is less to review, and review that scales. Automated review coverage already exceeds human review in the studied enterprise, which suggests the review pipeline is becoming infrastructure, not a meeting. And if token costs fall by a quarter to two-thirds when agents plan against abstractions instead of re-reading code, the cheapest token is the one never sent. m:…0994 m:…0799

The practical moves follow from the evidence. Write bug reports that suggest a fix and narrow the search space, because that is what predicts repair success. Plan agent work against program structure rather than raw text. And treat review as a system to build, not a chore to distribute, because the benchmark says four out of five hard tasks still defeat every frontier model at least once. The mandate delivered its throughput. The bill arrived in review, and it is due now. m:…0797 m:…9448