AI IN ENGINEERING · SEPTEMBER 1, 2026 · 4 MIN READ
Velocity Is Not Excellence
AI made writing code nearly free while the cost of verifying it stayed flat. The teams that win the agent era are the ones whose codebase gets better with every change.
By the Hyrax team
The way software gets written changed faster than the way it gets improved. AI coding agents now produce a large share of new production code. Pull requests arrive machine-written, machine-summarized, and machine-commented. GitHub folded Copilot code review into every Copilot plan on June 1, 2026, metered like the rest of Copilot. Cursor ships Bugbot alongside the editor. CodeRabbit reports more than two million connected repositories. When every diff gets an automated reviewer by default, automated review stops being a differentiator and becomes plumbing.
That is the commoditization of the diff. It is real progress, and it solved the wrong bottleneck.
The asymmetry#
Writing code became nearly free. Verifying it did not. Across 8.1 million pull requests measured by LinearB, AI-agent pull requests wait 5.3x longer for review pickup than unassisted ones, and only a third of them merge within 30 days. In a randomized controlled study by METR, experienced developers using AI tools were 19% slower on real tasks in mature repositories, while believing they had been 20% faster. And the people doing the verifying noticed: Stack Overflow's 2025 survey found 84% of developers using or planning to use AI tools while trust in their accuracy fell from 40% to 29% in a single year. The most common frustration, reported by 66% of developers, is AI output that is "almost right, but not quite."
What velocity does to a codebase#
A codebase is not a stack of diffs. It is an architecture, a set of conventions, a test suite, and a thousand decisions that either keep holding or quietly stop holding as code accumulates. Velocity stresses all of it. More merges per week means conventions drift faster, dead paths pile up faster, and the distance between what the system is and what the team believes it is grows faster. Review bots comment on the change in front of them, and nobody in that loop owns the system the changes accumulate into. Shipping keeps getting faster while confidence in the codebase falls behind.
Excellence is a different target#
A fast team merges more changes. An excellent team merges changes that make the next change easier, so that speed today makes speed tomorrow cheaper. In the agent era, the test of an engineering organization is whether the codebase gets better with every change.
Hyrax is built for that job. As the Codebase Architect, it reads the whole repository and learns how the system actually works: its structure, its conventions, its dependencies. From that context it finds what needs to change, writes the fix itself, and verifies the change against the team's own tests before anyone sees it. The output is a pull request in the normal GitHub workflow, approved by a human before it merges. Findings arrive with the exact file and line; fixes arrive as verified changes, and closed tickets follow.
None of this replaces the review bot, the security scanner, or the humans. Hyrax runs alongside CodeRabbit, Snyk, or SonarQube and completes the work those tools point at, one verified pull request at a time.
The advantage in the agent era goes to the teams whose codebase compounds. Hyrax is live today: start free with a repository audit, findings with file and line, and fixes submitted as pull requests.