6/22/2026
WebGPU feature detection was not enough to run small LLMs on phones
Technical analysis of running small LLMs in browser on phones using WebGPU, testing feature detection and buffer limitations.
6/22/2026
Fugu: Learn to assemble, route, and coordinate expert agents [pdf]
Fugu research paper on assembling, routing, and coordinating expert AI agents for complex multi-step tasks.
6/22/2026
Ask HN: What will AI coding look like when today's CS freshmen graduate?
Discussion about future of AI coding tools and LLM capabilities for developers over next four years.
6/22/2026
Sakana Fugu
Sakana Fugu API orchestrates multiple AI models dynamically for complex multi-step tasks without vendor lock-in.
6/22/2026
Visual Thought Simulator: Tiny demo showing why LLMs have no internal imagery
Demo exploring why LLMs lack internal imagery, proposing alternative systems that generate and store visual scenes using phonetic methods.
6/22/2026
Show HN: Woltspace – a lodge for your coding agents
Woltspace is a containerized sandbox platform for interacting with coding agents remotely via Telegram or Slack, supporting Claude Code with persistent memory.
6/22/2026
GitHub Banned All CI for Our (OSS) Org Because of Bad Drive-By Contributors
Open source organization's CI/CD disabled by GitHub due to crypto-mining activity by drive-by contributors.
6/22/2026
Give your sandboxed agents API keys they can't read
Developer tool that securely provides API credentials to sandboxed AI agents without exposing real values.
6/22/2026
Promptblock – detect prompt injections in GitHub issues
GitHub App detecting prompt injection attempts in issues and comments targeting AI agents, including hidden HTML payloads.
6/22/2026
Show HN: Phileas – Local-first memory that remembers you
Local-first memory layer enabling LLMs and agents to read/write persistent context across sessions via Model Context Protocol.
6/22/2026
Codex-maxxing for long-running work
Whitepaper on using Codex as persistent workspace for long-running projects, managing context and complex workflows beyond single prompts.
6/21/2026
Bean – a portable convergence gate for agent work
Bean is a runtime tool that prevents AI agents from declaring tasks complete until claims are verified, conflicts resolved, and questions documented. Maintains typed claim ledger and runs verification compiler.
6/21/2026
Show HN: ANMA, boundary contracts for cheaper AI coding agents
ANMA: YAML-based boundary contracts for enforcing architecture rules in cheaper AI coding agents, with benchmarks showing 100% compliance vs 32% without.
6/21/2026
Show HN: PeekAI – Local-first observability for Python AI agents
PeekAI provides local-first observability and debugging tools for Python AI agents, enabling developers to monitor agent behavior.
6/21/2026
Show HN: Lupen – an itemized, verified receipt for Claude Code and Codex spend
macOS tool that itemizes and verifies AI coding spend from Claude Code and Codex logs with granular breakdowns.
6/21/2026
China's Z.ai open-sourced a frontier coding model as Washington bans it rival
Z.ai open-sourced a frontier coding model amid US regulatory restrictions on rival systems.
6/21/2026
Samsung Electronics brings ChatGPT and Codex to employees
Samsung Electronics deploys ChatGPT Enterprise and Codex to employees globally, marking one of OpenAI's largest enterprise deployments.
6/21/2026
Good results fine tuning a local LLM like Qwen 3:0.6B to categorize questions
Personal project fine-tuning Qwen 3:0.6B for household question categorization using RAG with metadata-aware vector search.
6/21/2026
Show HN: Real-Life Deception Detection Without Uploading Video
Self-contained deception detection pipeline fusing text, visual, and audio features from courtroom video for truthfulness prediction.
6/21/2026
I Canceled My French Tutor and Built an LLM Tool That Does It Better
Developer built LLM-based French tutoring tool replacing paid human tutor, focusing on improved retention through interactive conversation and personalization.
6/21/2026
Compass – guardrails and a hard budget cap for AI coding agents
Compass: local-first config layer for AI coding agents with budget caps, guardrails, and safety features. Prevents unsafe commands and unverified code merges.
6/21/2026
I Gave an AI a Civilization to Run. It Built a Nuke – Launching CivBench
CivBench: benchmark for evaluating AI agents by tasking them to run civilizations in game environment.
6/21/2026
Show HN: Tunr – Expose your local server in 3 seconds
Tunr exposes local development servers to public internet in under 3 seconds with automatic HTTPS, alternative to ngrok built in Go.
6/21/2026
Show HN: EGC - MCP server that gives AI coding tools memory across sessions
EGC: MCP server giving AI coding agents persistent memory across sessions to retain decisions and learnings.
6/21/2026
Jonathan Blow on why LLMs cannot program [video]
Video: Jonathan Blow discusses fundamental limitations of LLMs in programming tasks.
6/21/2026
Apertus – Open Foundation Model for Sovereign AI
Apertus: open-weight foundation model (8B/70B) from Swiss AI Initiative. Open data, reproducible training, EU AI Act compliant.
6/21/2026
Securing the Future of AI Agents
Analysis of security requirements for increasingly capable AI agents in autonomous tasks like cybersecurity and R&D.
6/21/2026
Stop wasting tokens and re explaining your project between sessions
Recall: Local session logger that condenses Claude Code interactions into project summaries without external API calls or data transmission.
6/21/2026
Conduit – Self-hosted Bitcoin Lightning payments for AI agents
Conduit: self-hosted Bitcoin Lightning payment infrastructure for autonomous AI agents with spending guardrails.
6/21/2026
AuraText – a Windows overlay that brings AI into any text field URL
AuraText: Windows overlay for AI assistance in any text field with contextual prompt generation.
6/21/2026
The AI Conundrum: We are living in highly subsidized, interesting times
Analysis of LLM industry economics, subsidy structures, and timeline from research to production use in programming.
6/21/2026
You're probably paying GPT prices for tasks that don't need GPT
API platform for comparing and routing requests across multiple LLM providers with unified billing and model switching.
6/21/2026
AI Is Taking Over Hospitals
Study comparing ChatGPT diagnostic performance against physicians in medical settings.
6/21/2026
I am dreading our LLM-written incident report future
Opinion piece expressing concerns about LLM-generated incident reports becoming commonplace in tech operations.
6/21/2026
Show HN: Lelu – catch AI agents when they're manipulated at runtime
Authorization engine detecting when AI agents are manipulated via prompt injection or anomalous behavior at runtime.
6/21/2026
Show HN: Memory Magico – CLI based memory, wiki and deterministic sprints
Memory Magico: CLI tool for agent-focused memory, wiki, and deterministic sprint management with validation checkpoints.
6/21/2026
GreyFox – Free self-hosted AI proxy, token quotas, and local cache
Self-hosted AI traffic proxy with token quotas, response caching, and per-user limits via local Docker deployment.
6/21/2026
Someone Build a OS for Claude Code
ClaudiOS: Linux distro that boots directly into Claude Code AI agent with full system access, eliminating traditional OS desktop environment.
6/21/2026
A cheaper and safer agentic AI workflow
Case study of agentic AI coding using GLM-5.2 model service: $0.034 cost, 3 minutes, 2 mistakes vs 1 hour human attempt with 4 mistakes.
6/21/2026
Show HN: Pure Effect – Reproduce production bugs on your laptop without a DB
Developer tool enabling local reproduction of production bugs by isolating I/O from business logic without requiring production databases.
6/21/2026
Show HN: Vexyn – browser-only privacy tools with local AI (WebGPU)
Browser-only privacy tool suite running local AI models via WebGPU: JSON tools, Base64 encoding, QR codes, PDF/image processing without uploads.
6/21/2026
Two AI judges scored our agent's answer 0.85, but it never opened the file
Analysis of AI agent benchmark flaws: judge models cannot verify if agents actually accessed claimed information or checked absence claims.
6/21/2026
Show HN: A TypeScript Pokémon Crystal TUI for Agent Benchmarking
Pokémon Crystal text-based UI benchmark environment for testing AI agent capabilities in turn-based game scenarios.
6/21/2026
Show HN: Jacobi–IDE for Abaqus subroutine with analytical tests and AI diagnosis
Jacobi-IDE developer tool for Abaqus UMAT subroutines with AI-assisted diagnosis for computational mechanics simulation.
6/21/2026
AkaRouter – Flat per-call LLM API gateway (20x cheaper than Claude Max)
LLM API gateway offering Opus-equivalent models at 91% lower cost via pay-per-request pricing across frontier models.
6/21/2026
Show HN: Atizar-AI agents where the server runs approved actions, not the model
Agent framework where server executes pre-approved actions rather than delegating execution to language models for improved safety and control.
6/21/2026
Neuralwatt: Energy-based pricing for AI inference. Efficient prompts cost less
Energy-based pricing model for AI inference where prompt efficiency directly reduces computational cost to users.
6/21/2026
Googles specification (and tooling) for the LLM wiki
Google's Open Knowledge Format specification for universal, vendor-neutral knowledge representation as markdown with YAML. Includes agent proof-of-concept for automatic OKF bundle generation.
6/21/2026
Supervising AI Agents
Framework/methodology for oversight and control of autonomous AI agents. Title-only submission.
6/21/2026
Show HN: Askmaps.ai – Like ChatGPT with a Map
Geospatial chatbot tool combining ChatGPT-style interface with map visualization. Limited technical details.