Harvard study finds AI coding agents boost output but slow PR review; a personal repo test complicates it

·7 views

A Harvard working paper analyzing about 300 million work records from roughly 700 companies found that adopting AI coding agents raised pull requests by 23% and code output by 30%, while issue resolution rates did not change significantly. The bottleneck was review, as merge times rose 49% and revision requests nearly doubled. A developer tested this on two side projects started with Claude, finding that commits with AI co-author trailers were more often fix or revert commits than those without. However, that gap largely disappeared after accounting for commit title habits, the uneven timing of the comparison group, and commit size differences. The author concludes the two repositories alone cannot show whether AI-assisted commits are revised more or less often.

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)

Decoupling Laravel Telegram Webhooks with Queues and Redis Locks to Prevent Timeouts

Synchronous processing in a Laravel Telegram bot can exceed Telegram's webhook timeout when traffic spikes or upstream APIs slow down. Telegram then retries the same update repeatedly, which can cause duplicate orders, repeated charges, and spam notifications. The article recommends acknowledging webhooks immediately, pushing raw payloads to background queue workers, and using atomic locks keyed on the update_id to ensure each update is processed only once. It also stresses verifying the X-Telegram-Bot-Api-Secret-Token header with a dedicated middleware using hash_equals, so forged requests are rejected at the HTTP boundary.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer's Four Scripts Automate Stock Asset Packaging on a 2 GB Server

A developer describes how a rejected Adobe Stock upload, caused by an incorrect CSV header, led them to automate metadata creation. Their make_metadata.py script generates the exact header format each platform requires, while a tracker script prevents duplicate submissions after accidental re-runs. Additional tools convert Markdown to styled PDFs and bundle products with generated cover images and thumbnails. The whole pipeline was built to run within the limits of a single-core server with 2 GB of RAM.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

CodeSmith's Tool Architecture: Three Layers Funnel Dozens of Tools Into One Slot

CodeSmith, at version 0.5.0, offers roughly fifty built-in tools plus extension tools from MCP servers, dynamic libraries, and subagents, all plugged into a single interface. The design is organized in three layers: a minimal, host-agnostic framework Tool contract at the bottom, a production ToolSpec that carries most of the functionality, and a ToolDispatcher seam that keeps the engine core independent of the TUI. A central choice is that approval requirements are derived from declared capabilities, so tools that execute code require approval, tools that write files are suggested for approval, and others run automatically. The article notes that the dispatcher exists because the host context carries many terminal-specific services that cannot live in the runtime crate. Together these layers let the engine work against an interface rather than concrete host implementations.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

17,856 Publicly Reachable MLflow Servers Found in ZoomEye Index

A ZoomEye scan on 2026-10-12 returned 17,856 reachable MLflow instances, an experiment tracking and model registry platform now widely used in production. Such servers often hold datasets, model weights, evaluation outputs, and sometimes credentials logged as parameters. Many were deployed for research on shared infrastructure, alongside training environments, or with default settings that lack authentication. Without authentication, an exposed server can allow run creation, parameter logging, and artifact retrieval, potentially reaching the underlying object store. The count is likely a lower bound, since servers behind proxies without recognizable interfaces are not captured.

0 comments Read more at DEV Community