In developer tools, context is the whole product

Context beats capability

A model with a perfect understanding of the open file and no knowledge of your architecture will produce code that is locally correct and globally wrong. That failure is more expensive than a syntax error, because it passes review.

This is why repository indexing, retrieval quality and long-context handling have become the real battleground. Evaluate on your largest and ugliest module, not on a greenfield example, because the greenfield example is what every vendor optimised for.

Locally correct and globally wrong is the expensive failure. It passes review.

The review bottleneck

Teams that adopted agentic coding tools aggressively report the same second-order problem: generation stopped being the constraint and review became one. A thousand-line machine-authored diff takes longer to review properly than it would have taken to write, and the temptation to approve it on vibes is enormous.

The tools that will win the next cycle are the ones producing small, well-scoped, explainable changes rather than the ones producing the most code. Judge candidates on the size and legibility of their diffs, not on how much they can do unattended.