forkd spawns 100 microVMs in 101ms, ART trains multi-step agents

Sunday, 24 May 2026

Forkd spawns 100 KVM-isolated microVMs in ~101ms from a warmed parent, with 150ms BRANCH operations that snapshot mid-execution state for agent fan-out scenarios.

The security side heats up with a Laravel Lang supply chain attack affecting hundreds of historical Packagist releases and.

📊presenton offers AI presentation

GitHub

Open-source AI presentation generator with multi-provider support, desktop apps, PPTX export, and full self-hosting. Alternative to Gamma and Beautiful AI.

Deep Dive

TypeScriptSelf-hostingOpenAI

📈Devs now 20% faster with AI

Updated study shows experienced devs are now ~20% faster with AI tools, reversing 2025 findings that showed 20% slower performance despite perceived gains.

Lively ThreadPerformance

📦llama.cpp adds NVFP4 tensors

github·2 min read

Latest llama.cpp release adds NVFP4 MTP scale tensors with Qwen3.5 MTP tensor linking and aligned nullptr handling across platforms.

Takeaway

The NVFP4 tensor support expands quantisation options for local inference. Check if our models can benefit from these new precision formats, especially for memory-constrained deployments.

C++GGUFQuantisationInference

🚨Laravel Lang supply chain attack

snyk

Hundreds of historical Laravel Lang Packagist releases were republished with malicious code targeting credential theft and secret exfiltration.

Bigger Picture

MCP creates security attack surface

The supply chain discussion highlights a blind spot: MCP servers can execute arbitrary tools, but we lack signing or sandboxing standards. As piia-engram and other MCP tools proliferate, trust verification becomes critical.

Under The Radar

PHPSecurity

⏱️Rust selfbot detector uses timing

Discord selfbot detector built in Rust that uses timing physics patterns instead of machine learning approaches to identify automated behaviour.

RustDiscordSecurity

⚡forkd spawns 100 microVMs in 101ms

GitHub

Fork 100 KVM-isolated microVMs from a warmed parent in ~101ms using copy-on-write memory. BRANCH snapshots mid-execution state in ~150ms for agent fan-out.

Bigger Picture

Copy-on-write for agent fan-out

The real value in forkd isn't just speed but the BRANCH operation. Most agent frameworks spawn fresh instances for parallel thinking, losing accumulated state. This lets agents fork mid-reasoning, inheriting prior context while exploring different paths.

Top Voted

RustPerformance

Yesterday's Sentiment/Energised

Agents Get Faster, Security Gets Serious

Strong momentum in agent infrastructure with forkd delivering sub-second microVM spawning and ART bringing RL training to multi-step agents. The Laravel supply chain attack serves as a wake-up call, whilst productivity studies suggest devs are finally getting faster with AI tools.

🎯Stopping agents from code

Discussion on preventing coding agents from modifying unrelated parts of codebases, with practical strategies for scoping agent access and maintaining focus.

AgentsCursor

Learn/Multiple Mentions

What is MCP in AI development?

Model Context Protocol (MCP) standardises how AI tools share data and capabilities across different platforms and services. It appears in piia-engram's cross-tool memory, GrokSearch-rs web search, and Higress gateway hosting. For devs, MCP means AI tools can finally talk to each other without custom integrations.

InteroperabilityStandards

🧠piia-engram stores cross-tool

GitHub

Local-first memory system that shares our identity, preferences, and lessons across Claude Code, Cursor, Codex via MCP. No cloud, Apache 2.0.

Trending

PythonMCPLocal AI

🧠ART trains multi-step agents

GitHub

Agent Reinforcement Trainer uses GRPO to train multi-step agents on real-world tasks. Supports Qwen, Llama, and GPT models with serverless infrastructure.

Trending

PythonReinforcement LearningAgents

🔍GrokSearch-rs adds web search MCP

GitHub

Rust MCP server for Grok web search with Tavily-backed source retrieval, xAI native API support, and Firecrawl fallback for full-text extraction.

Trending

RustMCPGrok

⚙️triton-llm-kernel-lab ships GPU

GitHub

Small GPU kernel lab for LLM inference primitives using Python, Triton, PyTorch, and CUDA. Includes row-wise softmax, FP16 GEMM, and FlashAttention kernels.

Trending

PythonCUDAPyTorch

🛡️CyberStrikeAI ships native testing

GitHub

AI-native security testing platform integrating 100+ security tools with intelligent orchestration, role-based testing, and built-in C2 framework.

Bigger Picture

Reinforcement learning finally practical

Previous agent RL frameworks required significant infrastructure setup. ART's serverless approach removes that friction, making it feasible to train task-specific agents without managing GPU clusters. This could shift agent development from prompt engineering to reward engineering.

Trending

GoSecurityMCP

🛠️GSD Pi ships spec-driven agents

GitHub

Local-first coding agent for planning, implementing, verifying, and tracking project work from the command line with worktree-aware Git automation.

Trending

TypeScriptAgentsCLI

📚AI audited 80-component library

Developer used AI to audit documentation for an 80-component React library, sharing what automated documentation review caught and where it failed.

ReactAI Workflows

🔌new-api unifies AI model access

GitHub

Unified AI model hub converting various LLMs into OpenAI-compatible format with centralized gateway and asset management.

Deep Dive

GoOpenAI

🔐Agent supply chain security gaps

Discussion highlighting security vulnerabilities in AI agent supply chains and the lack of established security practices for agent-to-agent tool sharing.

SecurityAgents

🛡️754 cybersecurity skills

GitHub

Structured cybersecurity skills library mapped to MITRE ATT&CK, NIST CSF 2.0, ATLAS, D3FEND, and AI RMF frameworks. Works with Claude Code and major platforms.

Trending

Python

🔧DSPy vs hand-written prompts

Programming prompts with DSPy instead of hand-writing them, treating prompt engineering as a software development discipline with systematic optimisation.

Dev Tools

🌐Higress gains AI gateway features

GitHub

Cloud-native API gateway with AI-specific capabilities, MCP server hosting, and unified management for LLM and MCP APIs. Built on Istio and Envoy.

Trending

GoMCPKubernetes

Learn/Core Concept

How does quantisation reduce memory?

Quantisation reduces neural network precision from 32-bit floats to 8-bit or 4-bit integers, cutting memory usage by 75-87% with minimal accuracy loss. It converts high-precision weights into smaller representations using statistical techniques like calibration datasets. Today's llama.cpp NVFP4 tensors exemplify this, enabling large models to run on consumer hardware where they'd otherwise fail.

CompressionInference

Read online