AWS Lambda now runs 90-minute functions

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

⚡ AWS Lambda now runs 90-minute functions

🦀 Cloudflare saves 100TB of RAM with Rust

🎮 Unity ships official AI coding plugins

💧 Bun's Rust rewrite eliminates numerous memory leaks

🤖 Google's Agent Kit comes to Kotlin

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

⚡ AWS Lambda now runs 90-minute functions LINK

AWS Lambda functions running on Managed Instances can now execute for up to 90 minutes, a sixfold jump from the old 15-minute cap, though traditional synchronous requests keep the shorter limit.

The longer runtime suits media processing, ETL, AI inference, and large file transfers, but demands care with network connections and temporary credentials staying valid for the full duration.

Because Lambda still doesn't guarantee exactly-once processing and the retry window widens, AWS urges developers to build idempotency using Powertools for operations like payments or database writes; the change works in all regions where Managed Instances exist.

🦀 Cloudflare saves 100TB of RAM with Rust LINK

Cloudflare reclaimed over 100TB of RAM globally by tuning the consistent-hashing algorithm inside its Pingora-based Backend Router, whose pingora-ketama structures were consuming far more memory than the load-balancing service actually needed.

One fix packed the hash-and-index struct into a raw byte array with getters, sidestepping Rust's alignment rules that had inflated each entry to eight bytes, cutting consistent-hashing memory by 25% without resorting to `#[repr(packed)]`.

A deeper derivation of the error margin showed the team could generate 90% fewer hashes per server with no meaningful loss in distribution, and they rolled the new ring out gradually to avoid invalidating cached content and flooding origins.

🎮 Unity ships official AI coding plugins LINK

Unity has shipped official plugins that connect Claude Code and OpenAI's Codex to the engine, giving those coding agents skills authored and maintained by Unity's own teams instead of scraped forum posts and tutorials.

The Codex plugin launches with 31 skills spanning UI, 2D graphics, the URP render pipeline, audio, navigation, physics, in-app purchases, multiplayer, and localization, including setup for new projects and migration of older ones to URP.

Both plugins require Unity 6 or later; Codex installs with a single click from its plugin directory, while the Claude Code version installs via npm, aiming to stop agents from generating code targeting outdated engine versions.

💧 Bun's Rust rewrite eliminates numerous memory leaks LINK

Bun's runtime, bundler, and package manager have been rewritten from Zig to Rust, shipping in v1.4.0 and clearing 128 longstanding bugs by leaning on Rust's borrow checker and RAII cleanup to kill memory-safety faults.

Creator Jarred Sumner ran the AI-assisted port in four months using pre-release Claude Fable 5 across roughly 50 workflows and 64 parallel instances, validating against a TypeScript test suite of over a million assertions at a cost of $165,000 in tokens.

The Rust build stabilized a 2,000-operation bundling test at 609 MB instead of climbing past 6.7 GB and lifted HTTP throughput 2-5%, though 19 semantic regressions surfaced and Zig creator Andrew Kelley criticized merging a million lines of unreviewed transpiled code.

🤖 Google's Agent Kit comes to Kotlin LINK

Google has released the Agent Development Kit for Kotlin 1.0, a production-ready framework for building AI agents across Kotlin, Android, and JVM/server apps that reaches feature parity with the existing Python and Java versions.

Built on Kotlin Multiplatform, it declares tools via annotations that KSP turns into function schemas at compile time rather than runtime reflection, and supports hierarchical multi-agent delegation, session serialization, context compaction, and Java interop.

Android-native features let agents mix on-device and cloud inference through LiteRT-LM, ML Kit (beta), and Firebase AI Logic, while human-in-the-loop workflows require explicit confirmation before sensitive actions like a bank transfer run.

Other news you might like

• Alibaba Open Sources OpenCodeReview for AI-Assisted Code ReviewLINK

• Cloudflare Measures Origin TLS Preferences, Cutting Handshake Retries from 52% to 3.7%LINK

• SolidStart 2: Replaces Vinxi with a Vite 8 and Enters Maintenance as Its Role Winds DownLINK

• From bonkers to bananas; Doom!LINK

• Google’s open source EnvHarness lets AI agents train against environments that evolve with themLINK

• MariaDB 13 Expands Oracle Compatibility and Improves Developer Experience and ObservabilityLINK

🧰 Trending tools

Prelint: reviews AI-generated pull requests against your ADRs, docs, and past decisions, catching product drift before it reaches productionLINK

Switch: open-source workspace that brings people and AI agents into one shared room across Slack, Teams, Discord, or Mattermost, keeping context through handoffs.LINK

Mastra Factory: a TypeScript framework for building AI agents with workflows, memory, streaming, evals, tracing, and a Studio UI for testing.LINK

Kilo Code for iOS and Android: an open-source AI coding agent that sends push notifications when sessions need input, letting you dictate prompts and attach files remotely.LINK

Meridian: open-source, locally-run activity tracker that logs work with context, showing where time went and the blockers holding projects back.LINK

Tines 3B: builds and runs your team's automations, apps, and AI agents in one secure platform, keeping everything unified and easy to manageLINK

📚 Trending papers & reports

Scentree is a free Python tool that turns historical data into ready-made maps of possible futures for planning under uncertainty, automating a workflow that normally demands scarce statistical expertise.LINK

Small AI database agents mostly fail not from weak reasoning but from clumsy plumbing, and fixing five server bugs, changing no model, improved six models by up to ~21 cells out of 30.LINK

Computer-use agents get a training ground that makes them switch fluidly between clicking through apps and writing code to rebuild working software, improving them across five benchmarks and teaching them to visually double-check their own output.LINK

Code search engines get benchmarked at massive scale, revealing that today's AI code-matching tools lose precision and struggle on terabyte-sized codebases, with a cleanup-and-rewrite trick sharply improving the weaker ones.LINK

AI coding tools, studied through interviews at a large Swedish telecom, prove helpful but unreliable assistants, revealing a gap between management's productivity expectations and developers' caution about risk during company-wide roll-outs.LINK

See you tomorrow for a new dose of ☕️ Devshot!