Microsoft launches Copilot Code

Hi there, this is your daily ☕️ Devshot.

Microsoft launches Copilot Code LINK

• Microsoft has started rolling out Copilot Code, a natural-language coding agent inside its revamped Copilot app, letting workers describe an app, dashboard, tracker, or workflow and have the agent build it, distributed through the Frontier program.
• It uses Work IQ to pull context from emails, meetings, documents, chats, and connected data to fill in missing requirements, and users refine results in plain language or drop into the code directly with local files and a terminal.
• Apps run locally or publish to Microsoft's managed cloud for sharing across an org, and a project that outgrows one person can be wired to a GitHub repo as a draft pull request, handing off to developers via GitHub Copilot.

Anthropic open-sources Agent Skills LINK

• Anthropic has donated Agent Skills, the open-source project it released last December, to the Agentic AI Foundation, the Linux Foundation consortium that also hosts the Model Context Protocol.
• Agent Skills defines a standardized format for skills, bundles of task instructions, examples and automation scripts that improve an agent's output, so teams can share them without restructuring the underlying files.
• Because the data is organized consistently, agents can process it more easily, which the project says lowers the risk of hallucinations; the donation brings the foundation's maintained projects to 10.

Google Cloud launches AI agent host LINK

• Google Cloud introduced Agent Compute Engine, a managed host for AI agents that runs each sandbox as a microVM combining container-level hardware efficiency with hypervisor-grade isolation to keep compromised agents from spreading malware or reading other tenants' data.
• The service runs microVMs on stateless appliances backed by an immutable file system, verifies firmware with Google's Titan chips, applies CPU side-channel mitigations, and routes all agent traffic through local and external proxies that enforce policy-driven outbound brokering.
• Agent Compute Engine detects idle agents and shuts their sandboxes down, then quickly reactivates them, letting one physical node host orders of magnitude more live agent microVMs than conventional VMs, and will roll out in the coming months alongside an updated Agent Gateway.

AWS SageMaker adds custom training loops LINK

• AWS has made Training Sessions generally available in Amazon SageMaker AI, letting developers write their own post-training loops and scoring logic while AWS provisions and runs the GPUs, coordinates processors, recovers from failures and checkpoints progress.
• Developers direct individual operations, generating responses, computing updates, applying them, through API calls, keeping the model in GPU memory between steps; reinforcement-learning runs like GRPO need no separate inference endpoint, and billing is per processed token, pausing while your scoring code evaluates a batch.
• It ships in Northern Virginia, Oregon and Mumbai with a default quota of two concurrent fine-tuning sessions, supports gpt-oss-20b and other models via the Python SDK or SageMaker Studio, but publishes no cost or accuracy comparisons versus dedicated infrastructure.

Modular makes MAX engine source available LINK

• Modular has made the full source tree of its MAX inference engine publicly available with the MAX 26.7 release, covering the graph compiler, runtime, and device extensibility hooks under a new MAX license.
• Rather than open-source, it ships as "source available": the license permits building and distributing commercial products royalty-free, but bars AI-assisted re-implementations, commercial hosted inference services, and redistribution of modified source.
• Modular frames the distribution limits as preventing forks into incompatible MAX versions while still letting developers modify, experiment locally, and contribute improvements upstream to the high-performance server with its OpenAI-compatible LLM endpoint.

NgRx 22 adds resource extensions LINK

• NgRx 22 has shipped for Angular, adding resource extensions that customise Angular's resource API, dynamic deep signals for union types, a new typing helper, and ESLint v10 compatibility.
• Delivered through an experimental entry point, resource extensions wrap an existing resource and patch only its reset-while-reloading and throw-on-error-read behaviours, preserving the original type; four built-in extensions ship alongside a function registering defaults at app, route or component level.
• Upgrading needs Angular 22, TypeScript 6.0 and RxJS ^6.5 or ^7.5, with automated migrations for most breaking changes, but note Angular 22 changed resource timing, which one developer reported broke tests and was later logged as a breaking change.

Tutorials & Advice

> Software is slow because of bad design, not slow languages: Speed up software by redesigning it to do less work, not by rewriting it in a faster language, since most programs wait on input and output.

> I don't trust ON DELETE CASCADE: Replace automatic cascading deletes with soft deletes that flag a row and clean up children in small background batches, avoiding the giant locking transactions that cause outages.

> Building an Isolated Preview for AI-Generated HTML: Sandbox AI-generated HTML previews with an empty-then-minimal permission set, a strict content policy, and a throwaway origin, since each control blocks attacks the others miss.

> Presentation: AI is an Amplifier: Scale High Performance, Not Dysfunction: Treat AI as an amplifier that scales whatever already exists, so fix organizational systems first or it just accelerates dysfunction.

> Before you evaluate an AI SRE, give your existing agent context: Before buying an AI incident-response tool, feed your existing coding agent live cluster failures and recent changes, which cuts diagnosis time sharply.

Other news you might like

• Front-end observability at scale: How OpenReplay captures billions of daily events in real time with ClickHouse LINK

• How we computes bounding boxes for rotated shapes - tldraw: Infinite Canvas SDK for React LINK

• ObservabilityCON 2026: A guide to all the announcements from Grafana Labs LINK

• FFmpeg 9.1 Preps For Release With APV Vulkan Encoder, NVIDIA Optical Flow FRUC LINK

• Linux 7.4 Workqueue Changes Bring A "10 To 100 Times Faster" Optimization LINK

• Rust Style Guide LINK

• omnara-ai/omnara: The open-source managed agent platform. A self-hostable alternative to Claude Managed Agents and OpenAI's Agents API. LINK

• Ramp built an AI agent harness in six weeks — and LangChain’s CEO says it’s part of a bigger shift LINK

• Five AI agent harnesses took the same memory test. The results weren’t even close. LINK

Trending tools

Die With Me: a Mac app that tracks Claude and Codex usage, compares it with friends, and opens a shared chatroom during cooldowns. LINK

Kilo Code for JetBrains: open-source AI coding agent for JetBrains IDEs that builds and iterates across projects with flexible model choices. LINK

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

akta.​pro: provides private company data via API with event tracking, news, and entity resolution, helping finance and GTM teams source deals faster LINK

Onepin: a QA layer for AI voiceover that fixes mispronounced names, prices, and dates across 30+ voice engines without re-rendering takes. LINK

Beam.win: run agents and builds on a remote Mac mini or Studio, with each app window appearing locally over an encrypted connection LINK

Trending papers & reports

Diagram-building AI errors get sorted into eleven mistake types showing where models go wrong when turning written requirements into software design diagrams, and adding simple prompt rules measurably improved accuracy. LINK

Verified code speedups let an automated agent rewrite software to run faster while generating a mathematical proof the new version behaves identically, producing proven faster rewrites for 156 of 189 test functions. LINK

Checking a model's reasoning shows that re-asking an AI a slightly changed version of its own problem spots wrong answers, with flagged traces correct ~61% of the time versus a ~32% baseline, though it also wrongly refutes many right ones. LINK

Core-banking consolidation shows that bundling five separate banking services into one shared host delivered 2.3 times the read speed, used under half the memory, and cut compute costs 11-24%, while producing identical financial results. LINK

Autonomous system governance offers a blueprint that splits who proposes, decides, and executes an action in large organizations, so software agents can run hundreds of components safely with signed approvals and audit trails humans can no longer manage alone. LINK

See you tomorrow for a new dose of ☕️ Devshot!