Developer · Daily · June 11, 2026

the browser stack is shifting

Google Chrome is permanently dropping MV2 extensions and cutting off uBlock Origin bypasses, forcing ad blockers onto the more constrained MV3 API.[1]

Users are calling Chrome a RAM hog and slower than Firefox, while still keeping it around because some banking and government sites only behave properly in Chrome.[2]

A lot of developers are switching daily browsing and dev work to Firefox specifically to keep full uBlock Origin and other privacy extensions, treating Chrome as the "compat" browser.[3]

This is happening alongside a new high‑severity Chrome vuln, CVE‑2026‑11645, which adds more pressure to stay on top of updates.[4]

anthropic, bedrock, and data retention friction

AWS Bedrock now requires sharing data with Anthropic for Mythos and future models, with Anthropic retaining traffic for 30 days to detect misuse.[5]

For Fable specifically, Anthropic captures and stores all usage data for 30 days with no opt‑out, even for enterprise customers, and Microsoft has paused employee access to Claude Fable 5 and its Copilot integration for an internal review because of that.[6]

People building with sensitive data point out that PHI can’t legally go through Mythos/Bedrock flows without a BAA, which many don’t have.[7]

Bedrock’s Anthropic access is also reported to be 10–20× more expensive than using Anthropic’s own API, while it simultaneously exposes OpenAI models in the same place.[8]

On top of the policy and cost issues, Fable 5 is intentionally weak at frontier‑AI work, underperforms some rivals on benchmarks, and has poor cost/performance on document understanding compared to Gemini 3 Flash.[9]

ai coding tools: migrations vs everyday edits

The React compiler was ported from JavaScript/Flow to Rust and merged into main with around 123k lines generated over 2.5 months using AI tooling, which is the kind of large, reviewable migration these systems are actually good at.[10]

On more routine work, many developers say GitHub Copilot’s CLI experience is noticeably worse than Claude Code’s despite similar output quality, so they prefer Claude’s CLI for context‑heavy sessions.[11]

There’s a consistent thread of people spending more time fixing AI‑generated code than they save, with some internal teams complaining that AI raises the QA burden without a clear productivity win.[12]

Codex and Copilot are still described as the most reliable workhorses overall, with better performance‑to‑cost, cost‑visibility tools like CodexBar, and features like automatic /fast mode, while many say they simply don’t see colleagues adopting Claude Code as a primary assistant.[13]

Others intentionally mix Codex, Gemini, Claude Code, Cursor, and OpenCode to exploit specific strengths such as Claude’s longer‑context reasoning, Cursor’s token efficiency, and OpenCode being free across platforms, rather than betting on a single stack.[14]

agent frameworks are consolidating around langgraph and mcp

The langchain-community repo was archived, and people building real systems describe LangChain pipelines as messy, full of edge cases and broken tools, while LangChain itself moves toward more SaaS‑style offerings.[15]

In response, users are shifting to lighter approaches like Docker‑wrapped Claude code, Hermes, and SwarmClaw and complaining about having to rebuild agents when switching frameworks.[16]

LangGraph is now widely called the safest choice for production agents because it treats agents as stateful workflows with explicit execution semantics, retries, and failure recovery, which is already being used for multi‑tenant products like AI healthcare receptionists handling calls, SMS, and future EHR integrations.[17]

The new AgentHarness framework and its Apodex‑1.0 model collection target long‑horizon agent loops and plug directly into LangGraph, showing a stack coalescing around that model.[18]

Underneath all of this, MCP is turning into the default protocol for tools and memories across agents, but the July spec introduced breaking changes and current schema‑validation bugs and tool redundancy are already causing migration and governance pain.[19]

cloud infra costs, aws pain, and postgres scale‑out

A solo founder reported waking up to a hacked AWS account and a $42k bill despite running their actual app on a low‑cost VPS, while another example had Twitch spending $4,000 to serve 8,000 viewers on AWS, pushing people to look at cheaper options like $60/month Hetzner VMs.[20]

Security anxiety on AWS is up too, with 75 leaked AWS keys and even public "hack my AI instance" challenges making the rounds, and an open‑source 30‑point audit checklist gaining traction as a baseline for client accounts.[21]

AWS is trying to improve price/perf with new M9g/M9gd instances on Graviton5, which target more efficient general‑purpose compute.[22] On the data side, PgDog just raised $5.5M and is reportedly handling over 2M Postgres queries per second and more than 20TB of sharded data in production, letting teams scale Postgres horizontally across RDS and Aurora without app‑level sharding, though people complain about Kubernetes configuration being clunky.[23]

At the same time, plenty of serious apps are staying on encrypted SQLite—from the Finara personal finance app to embedded task queues and doc search servers—while Postgres used as a sidecar inside Splunk just produced a 9.8‑severity CVE, underscoring how much operational and security surface area comes with full client‑server DB stacks.[24]