TLDR Dev 2026-09-25

Almost right, but not quite” is developers' most-cited AI frustration in Stack Overflow's 2025 survey, and it stops being tolerable once AI agents handle production work. The shift reaches ordinary AI features too. Anything that chains model calls into multi-step work inherits the same failure classes.

Building Agentic Systems on AWS turns them into teachable patterns: on-demand technical guides and workshops on orchestration, state that survives sessions and restarts, failure recovery, human approval gates, and evaluation that tests whether a change made the agent worse. Start with the agent series in AWS Marketplace.

This article reflects on a rewrite of an Electron application as a progressive web app. It uses the migration to examine the technical and product trade-offs that appear when moving between desktop and web delivery.

This article shows a CSS-only approach to detect overlapping elements with anchor positioning and scroll-driven animations. It uses the result to adapt card layouts without JavaScript, pointing to a new way to build collision-aware interfaces.

CMU tested Jev, a judge that returns only a verdict and label probabilities, against 16 LLM and reward-model judges. It comes within three points of GPT-6 on preference and factuality at 277 times lower fees, though it fails on difficult correctness tasks.

These field notes describe an organization with fewer titles and meetings, more high-impact individual contributors, and agents with human owners responsible for their output. The trade-off is relentless pace, blurrier ownership, and a greater premium on judgment as everyone gains more leverage.

Personal AI needs to feel approachable, earn trust through well-timed permission requests, tackle avoided tasks, and suggest useful next steps. It uses Muse to show why proactive ideas and clear ways to manage ongoing work matter more than generic speed.

This essay uses the form of a funeral to reckon with AI making implementation work dramatically cheaper. It argues that engineers remain responsible for judgment, clear definitions of done, and passing hard-won taste on to newcomers.

Alchemy models cloud infrastructure and application logic as a single type-safe Effect program. It supports AWS and Cloudflare resources, while keeping local development, planning, deployment, testing, and CI in one workflow.

Step Code is a terminal coding agent that reads code, makes changes, and runs tests. It supports MCP servers, Agent Skills, plugins, multi-agent orchestration, and longer-running task delegation.

Search is a small, fast, quiet WebKit browser for macOS with an intentionally minimal interface. It uses the system WebKit engine, which keeps the app small and avoids bundling a second browser engine.

This post introduces a Parquet writer in DataHaskell/Dataframe. It centers on the writeParquet function, which takes a dataframe and writes it to Parquet format.

This experiment creates a fictional garage and sends several AI agents to interact with it. It explores what happens when agents encounter a small business on the open web.

Meta describes using confidential virtual machines and trusted execution environments to run personal-context workloads in the cloud. Remote attestation lets clients verify the code image before sending data, while the design aims to protect memory even from infrastructure operators.

Give them building blocks with Expo MCP. AI app builders like Replit, Emergent, and Anything rely on Expo to ship their customers' mobile apps. Cloud simulators coming soon! Join the waitlist

Braintrust describes Nitro, which separates storage and compute concurrency and made full-text trace searches more than twice as fast in a production A/B test.

Google details its first in-orbit TPU test and the research challenges around radiation, cooling, and satellite interconnects for space-based AI compute.

Docker outlines microVM-based Sandboxes, OCI-based Sandbox Kits, and Cloud Sandboxes as a way to constrain, package, and extend agent work.

Sentry explains how its bfcacheMetricsIntegration makes back/forward cache health measurable when ordinary metrics leave cache hits invisible.