JetBrains is rolling out AI capabilities for teams, adding shared context, reusable agentic workflows, org-level governance, and cost control across an organisation's AI usage.
Bigger Picture
Sandboxing isn't the whole security story
CubeSandbox nails fast, isolated boots, but the CNCF piece on agent sandboxing argues isolation alone still leaves egress and permission gaps open.
A widely discussed thread describes a user encountering someone else's chat content unexpectedly inside Claude, raising questions about session isolation in production LLM apps.
A self-taught dev argues that leaning on AI answers instead of reading docs quietly erodes our understanding of the systems we depend on, and pushes back towards deliberate
Takeaway
Worth a gut check on our own habits: if we can't explain a system without asking an assistant first, that's a skill gap worth closing before it costs us in an incident at 2am.
A widely discussed thread catalogues ChatGPT's repetitive phrasing habits and calls for OpenAI to fix its default writing style, with devs sharing workaround prompts.
A four-part series argues that observability for app, infrastructure, CI, and LLM layers each need a differently-shaped data model rather than one unified tracing scheme.
Takeaway
If our current dashboards mash LLM traces into the same shape as regular app logs, this is a useful nudge to rethink the schema before it becomes unreadable.
A community thread details how Qwen 3.6 27B falls apart specifically on agentic workflows despite decent benchmark scores, a reminder that leaderboard numbers don't capture
ai-job-search is a Claude Code workflow that scores job fit, drafts tailored CVs and cover letters, and preps interview answers using a structured, forkable skill pipeline.
A widely shared analysis argues the AI infrastructure buildout costs far more to run than current AI revenue covers, and questions who ends up absorbing that gap.
Hot TakeCost OptimisationLLM Ops
Learn/Multiple Mentions
Why give AI agents a sandbox at all?
A sandbox is an isolated environment where an agent can run code or commands without touching our real system, limiting damage if it misbehaves or gets hijacked. Tools like CubeSandbox boot isolated VMs in under 60ms, but CNCF warns isolation alone isn't full security.
The Government of Alberta has been using Claude Code with Opus and Sonnet models since 2025 to review its systems, find vulnerabilities, and fix them.
Takeaway
Decent proof point if we're pitching AI-assisted vulnerability scanning internally; worth checking whether Opus vs Sonnet split makes sense for our own security review workflow.
ThinkingCap-Qwen3.6-27B matches base Qwen3.6 accuracy while using roughly half the thinking tokens, a fine-tune aimed squarely at cutting inference cost for reasoning tasks.
QwenFine-TuningCost OptimisationInference
Yesterday's Sentiment/Energised
Agent Infrastructure Grows Up
Momentum is building around agent plumbing: CubeSandbox and TencentDB Agent Memory show sandboxing and memory maturing fast, while community pushback on Qwen 3.6's agentic reliability keeps expectations grounded.
A CNCF blog post argues that sandboxing an AI agent's execution environment isn't sufficient on its own, and outlines what additional controls agentic systems actually need.
Takeaway
Good read before we assume CubeSandbox or a similar isolated VM is enough for untrusted agent code; worth pairing sandboxing with proper egress and permission controls.
A wave of community comparisons of MTP speculative decoding and Qwen 3.6 variants shows devs actively benchmarking which local model setup gives the best speed-to-accuracy
OfficeCLI is a single-binary, open-source Office suite that lets AI agents read, edit, and automate Word, Excel, and PowerPoint files without needing Office installed.
A dev's test asking Gemma 4 12B (Q8_0 quantised) to write a single-file 3D bowling simulator in WebGL produced rough but surprisingly functional output, useful as an informal
OfficeCLI's SKILL.md pattern lets any AI agent read a single file and self-install the tool and its commands, a lightweight alternative to MCP servers for giving agents new
A dev built a live interpretability viewer for an open model after being unable to apply Anthropic's J-space research to Claude directly, since its weights are closed.
fable-harness applies an OODA loop, multi-party adversarial review, and model routing to Claude Code, aiming to make it behave like a disciplined engineer instead of a fast
Reuters reports Beijing is considering restrictions on overseas access to China's leading AI models, a move that could affect devs currently building on open Chinese model weights.
TencentCloud's CubeSandbox is an E2B-compatible sandbox for AI agents built on RustVMM and KVM, claiming hardware-isolated boots in under 60ms with snapshot and rollback support.
Bigger Picture
Benchmarks versus real agentic behaviour
ThinkingCap-Qwen3.6-27B claims equal accuracy with fewer tokens, but the same base model's struggles at agentic tasks show benchmark parity doesn't guarantee tool-use reliability.
karakeep is a self-hostable bookmark manager for links, notes, and images that adds AI-based automatic tagging and full text search on top of our own saved content.
wechat-article-search is a Claude and Codex agent skill that searches WeChat public account articles by keyword via Sogou, returning structured JSON with no API key needed.
The.NET team published a curated set of skills and custom agents covering Entity Framework, MSBuild, NuGet, Blazor, MAUI, and test migration for coding agents working in.NET
oh-my-openagent is an agent harness aimed at complex codebases, built to work with Codex and OpenCode setups for what it calls tokenmaxxing on hard engineering tasks.
boring computers gives an AI agent an on-demand Firecracker microVM with a full Linux desktop, browser, terminal, and coding agents preinstalled, self-hosted with our own keys.
claude-keysmith installs and manages import blocks in CLAUDE.md files, making it easier to compose local instruction files without hand-editing config every time.
TencentDB Agent Memory offers fully local, layered long-term memory for agents, claiming big token cuts and accuracy gains over flat vector storage in its own benchmarks.
Bigger Picture
Skill files as the new lightweight MCP
OfficeCLI's self-installing skill file and the dotnet/skills repo both lean towards packaged skill cards over full MCP servers, a lighter pattern worth adopting for narrow internal tools.
bw24 is a from-scratch Rust and CUDA inference engine tuned for a single RTX 5090, claiming to beat llama.cpp on generation speed for Qwen3.5/3.6 dense (9B/27B, NVFP4 hybrid) and Qwen3.6-35B-A3B MoE models via speculative decoding.
OpenMed offers over 1,000 local-first medical models for clinical NER and HIPAA PII de-identification running entirely on-device via Apple MLX and Python, no cloud calls needed.
Cognitive Core Skills is an industry-neutral taxonomy of the operating capabilities an agent needs: perception, memory, reasoning, planning, action, verification, learning, and governance.
Ghost-Downloader-3 is a cross-platform, multi-protocol concurrent downloader built with Python and Qt, with AI features layered onto its fluent-design interface.
TrendingPythonOpen Source
Learn/Core Concept
How does speculative decoding speed up LLMs?
A small draft model guesses several tokens ahead, then the big model verifies them in one pass, accepting matches and only recomputing where it disagrees. It cuts latency without changing output quality, as seen in bw24's RTX 5090 engine and community tests of MTP variants.
Rust
TypeScript
JavaScript
Go
Python