Hud and ClickHouse aim runtime data at the part of AI coding nobody automated: the review

ClickHouse and Hud have connected Hud's Runtime Code Sensor to ClickStack, ClickHouse's open-source observability stack, so that production behavior can be used to assess code changes before deployment, verify them after release, and investigate or reverse them when behavior deviates ^[3][5]. It matters because the constraint in AI-assisted engineering has moved downstream from authoring: Hud says AI now generates or assists 42% of the code developers ship, and expects that to reach 65% by 2027 ^[1][2].

That is 23 percentage points more machine-written code inside roughly two years, a relative increase of about 55% ^[14]. Nobody is planning a matching increase in reviewers. The publication's framing is the honest one: generation is getting faster while reviewing impact, validating releases and responding to unexpected behavior still require context from production ^[16].

The division of labour between the two products is the substantive part. ClickStack supplies breadth, helping teams identify the service, deployment or endpoint associated with an issue ^[6]. Hud sits closer to the code, tracking function-level behavior and tying production activity back to the functions and changes responsible for it ^[7]. The join is done by a coding agent using shared trace IDs, so an issue surfaced in ClickStack can be followed into code-level context in Hud ^[8].

What teams get, according to the companies, is a pipeline with two decision points. Before deployment, changes are scored against real-time data on how the affected code behaves in production, with higher-risk changes held for extra review and safer ones moved faster or merged automatically ^[11]. After deployment, the integration covers release verification, automatic reversion on regression, automated detection and investigation, and agentic workflows that open pull requests against the underlying code issue ^[12]. The stated target is small drift before it compounds: a query that slows, a function that misbehaves under a particular workload, a code path that starts consuming more resources ^[13].

The conceptual shift is worth naming. Mike Shi, ClickHouse's Head of Observability, says users already trust ClickHouse to store and query their OpenTelemetry data at scale, and that the change underway is from watching systems and investigating incidents to using that data for day-to-day engineering decisions ^[15][10]. Hud CTO May Walter puts it as production context being a precondition for shipping AI-generated code with confidence ^[9]. In other words, telemetry stops being a forensic archive and becomes an input to merge policy.

Which is where the announcement runs out of evidence. Automatic merge and automatic revert are both bets on a signal: a weak signal lets regressions through, an oversensitive one turns the deployment pipeline into a slot machine. The published material includes no accuracy or false-positive figures, no general-availability date, no pricing, and no measured results from monday.com, which is named as an early adopter of the combined stack ^[4][17]. The 42% and 65% figures are Hud's own, and the source does not say how they were derived ^[1][2][17].

Worth watching: whether early adopters run the pre-deployment score as a hard gate or as advisory metadata, since that distinction is the whole claim; whether automatic reversion is enabled in production or left off after the first bad rollback; and whether either company publishes regression-catch rates rather than workflow diagrams. Until then this is a credible architecture for closing the loop, not proof that the loop closes.