Google sets Android memory limits for 2027

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

📱 Google sets Android memory limits for 2027

🤖 OpenAI builds always-on Codex agent

📌 AI model burns 70x more tokens

🍍 Cloudflare's cache tweak slashes memory waste

🐧 Ubuntu 26.04.1 LTS opens 24.04 upgrades

⚙️ Kubernetes drops legacy kube-dns and cgroup v1

Plus: 🎁 6 other news you might like, 🧰 6 tools, and 📚 5 papers.

📱 Google sets Android memory limits for 2027 LINK

Google is rolling out stricter Android memory rules that will throttle or force-close apps exceeding new thresholds, giving developers until February 2027 to shrink their memory footprint for low-RAM devices.

The policy targets three areas: dynamic memory usage (anonymous RSS plus swap), bitmap memory from images, and DEX code, with apps required to free memory once they move to the background or lose user interaction.

Developers must also cut app code by at least 25% using tools like R8 before publishing to the Play Store, a response to manufacturers dropping entry-level phones from 6-8GB down to 4GB amid the memory shortage.

🤖 OpenAI builds always-on Codex agent LINK

OpenAI is reportedly building a "Persistent mode" for Codex that keeps its coding agent running continuously until a user manually stops it, with code first surfacing in the command-line version of the tool.

The always-on mode adds a "reasoning effort" setting so developers can cap how much compute the agent spends overnight, curbing runaway token use while it works unattended on tasks.

A companion "Proactivity" feature would let Codex generate its own follow-up tasks without further prompting, keeping the agent working past the initial request, though its value hinges on how well it predicts what to do next.

📌 AI model burns 70x more tokens LINK

Benchmarks running the same model through different coding-agent harnesses found token use per solved task varying up to 70-fold, from ~3,500 for Aider's architect mode to 292,000 for OpenClaw, pointing to harness software rather than model behavior.

The spread traces to a "startup tax", each harness's system prompt, tool descriptions, and setup resent every turn, ranging from ~700 tokens for Aider to ~26,000 for OpenClaw, with floor times turn count predicting total tokens at R-squared 0.99.

Caching reordered the cost leaderboard: Claude Code drew only 1.5% of input from cache versus ~70% for Codex, so despite using fewer raw tokens it billed over four times as much per verified task, with hit rates depending on gateway and endpoint dialect.

🍍 Cloudflare's cache tweak slashes memory waste LINK

Cloudflare reworked how its Big Pineapple DNS platform stores over 250 billion cache entries, cutting per-entry memory by more than 50% and reclaiming roughly 100 terabytes of RAM across its fleet.

The team swapped `Vec` and `String` for `Box<[T]>` and `Box` to drop unused capacity fields, merged separate record sections into one list with `u16` offsets, and packed booleans into bitflags.

Storing records as a single wire-format `Box<[u8]>` buffer instead of boxed enum variants improved memory locality and let `A`, `AAAA`, and `TXT` copy directly, raising insert throughput 43% and lowering lookup latency 19%.

🐧 Ubuntu 26.04.1 LTS opens 24.04 upgrades LINK

Canonical shipped Ubuntu 26.04.1 LTS, the first point release rolling up every update since April into a fresh ISO, and it opens the official upgrade path from 24.04 LTS to 26.04 LTS.

Canonical held back 24.04 upgrades until this release so early 26.04 bugs could settle, giving stability-focused users a safer jump while impatient ones could already switch by tweaking settings.

The ISO now covers Raspberry Pi, Arm64, and RISC-V, but there's no new hardware enablement stack or features, only security updates, so existing 26.04 users can just run the update manager instead of re-imaging.

⚙️ Kubernetes drops legacy kube-dns and cgroup v1 LINK

Kubernetes v1.37, nicknamed Garhwal, retires the legacy kube-dns server in favor of CoreDNS and continues phasing out cgroup v1, part of a broader push to clear technical debt across 67 changes.

Clusters still running kube-dns must migrate to CoreDNS before v1.40 or stop working, while cgroup v1 nodes have failed to initialize since v1.35, leaving only a temporary workaround supported for a limited time.

The release also promotes the metrics.k8s.io API to GA after nine years in beta with no functional changes, and confirms kube-proxy will drop Linux IPVS support in v1.43 in favor of nftables.

Other news you might like

• Your AGENTS.md file doesn’t actually do anythingLINK

• Replit’s new default: Auto mode picks the best model for each taskLINK

• Spring News Roundup: First Milestone Releases for Boot, Framework, Data, Security, Modulith, BatchLINK

• Anthropic previews MHS standard for AI agents that operate machinesLINK

• Gemini Omni 1.1 Flash lets you build with more controlLINK

• Claude, Codex, and Hermes installed unowned code inside corporate networksLINK

🧰 Trending tools

Sim: build and deploy AI agents via a visual or code workspace, with 1,000+ integrations and support for all major LLM providers.LINK

Prelint: checks AI-generated pull requests against your ADRs, docs, and past decisions, catching product drift before it reaches production and causes problemsLINK

Humalike x Hermes: gives AI agents turn-taking, timing, and memory APIs so they know when to speak, wait, or interrupt in conversations, useful for game characters and roleplay.LINK

Coldtea.ai: automates QA and monitoring with AI agents that catch post-deployment bugs, flag issues, and generate actionable fix tasks for your teamLINK

Kane CLI: an AI testing tool that runs plain-English commands in a real Chrome browser, eliminating brittle selectors and scripts for developers and QA engineers.LINK

Openbase: helps developers choose npm and open-source packages by comparing popularity, maintenance activity, and reliability metrics alongside real user reviews.LINK

📚 Trending papers & reports

Scientific computing's AI shift gets a community roadmap of eight priorities, from trust and traceability to workforce training, aimed at keeping research infrastructure reliable and sustainable as AI takes on more of the discovery process.LINK

Multi-agent AI security builds guardrails directly into how autonomous AI systems are designed rather than bolting them on later, blocking attacks like prompt injection and agents secretly colluding by forcing every action to be verified.LINK

AI agent economics offers businesses a framework to judge whether each step an AI assistant takes is worth its cost in time, money, and computing, pinpointing five ways agents waste resources without adding value.LINK

Agent guardrails can make AI assistants give identical results every run on regulated finance and legal tasks, hitting perfect reproducibility and 100% success in three of four test cases once free-text planning is locked to a fixed format.LINK

Open-source satellite software gets its first big survey, mapping over 22,000 public projects into 43 purpose categories and 66 programming languages, showing a fast-growing but fragmented ecosystem for anyone building space tech.LINK

See you tomorrow for a new dose of ☕️ Devshot!