The Hospital Codebase: How CockroachDB Ran a Five-Month AI Bug Triage Experiment
How this was made Verified AI
Every Intellegix briefing is generated from that day's broadcast and run through automated checks before it publishes — with a human paged on any flag. Here is the trail for this edition.
A CockroachDB engineering blog post describing five months of treating the company's codebase like a hospital — bugs as patients, AI coding agents as doctors — drew 145 points and 58 comments, with engagement driven by results that surprised even the team running the experiment. The methodology assigned AI agents to bugs through a triage system modeled on hospital workflow: incoming bugs were classified by severity and complexity, dispatched to agents for diagnosis and attempted repair, and escalated to more capable agents or human engineers when fixes failed validation.
The metric most scrutinized in the comments was the regression rate — the frequency with which agent-generated fixes introduced new failures. After five months, the team reported that agent fixes produced regressions at a rate roughly comparable to junior engineer fixes on the same category of bugs, which they described as a success criterion they had not expected to meet. The critical caveat: the experiment deliberately bounded its scope to reproducible, well-characterized failures with clear test cases, excluding speculative or architectural problems from the agent queue.
Several Hacker News commenters argued the post is as much a story about testing discipline as about AI capability. CockroachDB's distributed SQL codebase is widely regarded as correctness-critical, and the hospital model only functions if automated quality signals exist to validate fixes. Teams without comprehensive test suites cannot run the same experiment because the diagnostic infrastructure is absent. The broader implication, drawn out across multiple comments, is that investment in testing now pays compound returns as AI coding tools mature.
Two related engineering posts reinforced the theme. A piece by Farid Zakaria described building a debugging tool and discovering that Nix's hermetic build model — which guarantees that debug symbols and source files match the exact binary being debugged — had solved a layer of manual debugger work he hadn't anticipated, effectively writing half the tool for him. Separately, documentation on Celld introduced deterministic simulation testing to a new audience: the technique, made prominent by FoundationDB, controls every source of non-determinism in a distributed system to allow exact scenario reproduction. A keyboard-centric terminal diff viewer called Diffle, featuring Language Server Protocol integration for semantic code understanding within diffs, rounded out a day of developer tooling coverage.