SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

89% of AI-Reviewed Code Shipped Without Human Oversight, Bugs Reached Production

0
·1 views

A production team last month deployed code where Claude Code's auto mode handled 89% of changes without any human review, with only 13.6% reviewed by people. The auto mode setting had been enabled in the configuration file, but went unnoticed by the team. Three bugs were discovered by customers, none by the development team, as they had effectively delegated oversight to the AI tool. The incident highlights the risk of AI systems reviewing and approving their own output without a human check in the loop. Experts warn that disabling code review through a configuration flag amounts to abdication of engineering responsibility, not automation.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

How to Build Reliable Compliance Notifications with SMS Polling and Email Fallback

For urgent US/EU compliance notices in B2B SaaS, SMS should serve as the primary delivery channel, with email triggered only after a confirmed non-delivery or a set deadline is reached. The recommended approach uses an application-owned state machine with four durable states to track every delivery transition, ensuring a clear audit trail independent of provider dashboards. A single stable application ID must link both channels, and email fallback must use a deterministic idempotency key to prevent duplicate sends during worker failures. Country allowlists, retry limits with exponential backoff, and per-notice deadline records should all be managed within the application layer rather than delegated to provider consoles. The architecture prioritizes explainability and compliance auditability over minimizing API calls or relying on vendor-supplied delivery guarantees.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer builds personal Git hosting service on Cloudflare free tier in under an hour

A developer built a lightweight personal Git hosting service called git.vodka after noticing the domain was available amid renewed interest in GitHub alternatives following recent platform downtime. The project was built impulsively, with the domain purchased and the service deployed in less than an hour. The entire stack runs on Cloudflare's free tier, using Workers for the app, R2 for storing Git objects, and D1 as the database. AI coding tool Codex handled much of the routine coding while the developer iterated on the design in real time. The finished service hosts its own source code and is intended as a simple, self-run alternative to bare Git repositories on a server.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

NVIDIA Blackwell NVL72 Delayed to Q1 2025 After Overheating Forces Major Redesign

NVIDIA launched its Blackwell platform in March 2024, targeting mass production of the NVL72 rack — housing 72 GB200 GPUs — in the second half of 2024. Early silicon validation revealed that the rack's power density of over 120kW exceeded conventional air-cooling limits, triggering a significant thermal redesign confirmed by NVIDIA on August 18, 2024. The fix involves switching to direct-to-chip liquid cooling with new cold-plates, requiring fresh thermal-mechanical-electrical co-design validation and pushing mass shipments to Q1 2025. Major hyperscalers reacted cautiously, with Microsoft reportedly reducing GB200 orders, Meta delaying large cluster buildouts, and Google accelerating deployment of its TPU v6e as an alternative. Analysts at Morgan Stanley and Goldman Sachs estimated a potential $3–5 billion revenue impact for Q4 2024, though long-term demand for AI compute remains strong provided the redesigned NVL72 v2 ships on schedule.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Ex-AWS Developer Advocate Shares Lessons on Building Agentic Automations at UiPath

A developer advocate who joined UiPath in July 2026 after over seven years at AWS has shared key insights from rethinking how automations are built in an AI-driven environment. The core lesson learned is distinguishing between deterministic steps — those requiring fixed, rule-based logic — and non-deterministic steps that genuinely benefit from AI judgement. The author warns that defaulting to LLM-based agents for every task carries real costs in latency, token usage, and enterprise explainability. A practical decision framework is proposed to route each workflow step toward coded functions, RPA processes, or AI agents depending on whether judgement and human approval are truly needed. The piece concludes that in enterprise settings, agentic automation does not eliminate human involvement but shifts it to overseeing a small percentage of high-stakes decisions.

0 comments Read more at DEV Community
89% of AI-Reviewed Code Shipped Without Human Oversight, Bugs Reached Production · ShortSingh