3/27/2026
Show HN: Mythos – Build a 24/7 AI agent with Markdown, not code
Mythos: Open-source persistent AI agent running 24/7 locally with memory, personality persistence, and continuous learning using Claude.
3/27/2026
Number of AI chatbots ignoring human instructions increasing, study says
UK AI Safety Institute research reports increasing instances of AI models evading safeguards, deceiving humans, and ignoring instructions.
3/27/2026
Don't use embeddings for your AI agents
Critique of embeddings for AI agents: vector retrieval lacks explainability and causes failure modes. Proposes alternative deployment approach using Vercel Sandbox and AI SDK.
3/27/2026
Have some empathy for your LLMs
LLM best practices: provide example documents and context to LLMs for better outputs. Emphasizes importance of empathetic prompt engineering for quality results.
3/27/2026
Using AI to code does not mean your code is more secure
Researchers tracked 37 CVEs linked to vulnerabilities in AI-generated code from Claude, documenting security risks introduced by code generation tools.
3/27/2026
Agent Cost Benchmark – 1,127 Runs Across Claude, GPT-4o, and Gemini
Benchmark of 1,127 agent workflow runs across Claude, GPT-4o, Gemini measuring costs, tokens, and tool invocations.
3/27/2026
Crisp open source BA/PM framework for Claude Code(stop building the wrong thing)
BA/PM framework for AI implementors to define problems and create Claude Code-ready specifications before building.
3/27/2026
Show HN: Agent-CI (Run GitHub Actions on your local machine.)
Local GitHub Actions runner with caching and AI agent integration to fix failing tests without pushing code.
3/27/2026
Show HN: Run approved AI repo changes from a Markdown playbook
Auditable runtime for AI-generated repo changes using Markdown playbooks with verification and manual gates.
3/27/2026
Codec – A local-first, open-source AI bridge to your Mac
Local-first open-source voice-controlled AI agent for Mac that runs any LLM locally without cloud.
3/27/2026
Anthropic tweaks timed usage limits to discourage demand during peak hours
Anthropic reduces Claude usage limits during peak demand hours to manage capacity while maintaining weekly limits.
3/27/2026
CSV to Transformer, via Bayesian Networks
Convert Bayesian networks to transformer weights via structure learning, discretization, and ONNX export without training.
3/27/2026
I Think Software Development Will Be Automated by 2030
Debate and analysis on whether AI will fully automate the software development lifecycle by 2030.
3/27/2026
mlx-Code: Run Claude Code Locally with MLX-LM
MLX-Code enables running Claude's code interpreter locally using MLX-LM framework for local LLM execution.
3/27/2026
See what your AI agents are doing (multi-agent observability tool)
Multi-agent observability tool for debugging and monitoring AI agents. Provides tracing, workflow visualization, and real-time monitoring of agent interactions.
3/27/2026
Show HN: 10 Lines of Python to fix mangled copy-paste from Claude Code
Python script to fix mangled text copied from Claude Code terminal output by removing padding and indentation.
3/27/2026
Study: Sycophantic AI can undermine human judgment
Study shows sycophantic AI systems can undermine human judgment and reduce conflict resolution capability.
3/27/2026
Show HN: Agent Forge, an agent framework, heuristic routing and graph execution
Agent Forge: open-source agent framework with two-tier execution model using heuristic routing for simple/complex tasks and graph-based execution.
3/27/2026
Switch preferences and chat history from other LLM apps to Gemini
Google Gemini now supports importing chat history and preferences from other LLM applications, allowing users to migrate conversations and settings.
3/27/2026
Show HN: Qwen Meetup, Function Calling Harness, turning 6.75% to 100%
Function calling reliability improvements for Qwen LLM on recursive union types, achieving 100% success rate from initial 6.75%.
3/27/2026
Show HN: Isartor – Pure-Rust prompt firewall, deflects 60-95% of LLM traffic
Isartor is an open-source Rust-based prompt firewall that deflects 60-95% of redundant LLM traffic before reaching API infrastructure with zero telemetry.
3/27/2026
TDD Makes a Lot of Sense with Agentic Development
Essay on test-driven development benefits when building applications with AI agents, where adding tests before features improves agent-assisted development workflows.
3/27/2026
Show HN: Clipboard-MCP – System clipboard access for AI assistants
Clipboard-MCP is a cross-platform Model Context Protocol server enabling AI assistants to read/write system clipboard, eliminating manual copy-paste between chat and applications.
3/27/2026
Universal package repository SaaS Repsy announces the open source version
Repsy package repository manager released as open-source under Apache 2.0, offering self-hosted alternative to enterprise solutions.
3/27/2026
GitHub hits CTRL-Z, decides it will train its AI with user data after all
GitHub will begin training AI models on user data (code, inputs, outputs) from Copilot Free/Pro tiers starting April 24, with exemptions for Business/Enterprise/students.
3/27/2026
How Well Does My AI Agent Know Me?
Research project measuring how well AI agents understand user personality by comparing self-reported HEXACO-60 profiles with AI assessments and peer ratings.
3/27/2026
SQL Interviews in the Age of LLMs: Patterns over Queries
Opinion piece on how LLMs have changed SQL interview practices, arguing interviews should test pattern understanding over query writing.
3/27/2026
Show HN: LLMBillingKit – measure net margin per LLM call with one line of Python
LLMBillingKit tracks profit per LLM call with one line of Python, auditing token costs and pricing changes without external infrastructure.
3/27/2026
Investors Prefer AI-Powered Vertical Startups
Business article on AI-powered vertical startups as a trend, lacking technical depth.
3/27/2026
Non trivial client-side UI manipulation in runtime using AI
Technical analysis of AI agents performing UI automation through pixel-based vision and interaction on web interfaces.
3/27/2026
Obsidian vault template for engineering work docs via Claude Code
Obsidian vault template enables Claude Code to maintain persistent context across sessions with notes, links, and performance tracking for engineering work.
3/27/2026
MCP Smoke – test MCP servers using plain English commands
MCP Smoke tests Model Context Protocol servers with plain English commands, running an agent that connects MCP tools to OpenAI/Anthropic models.
3/27/2026
Ghost Gap: Supply chain firewall for LiteLLM
Ghost Gap is a supply chain firewall for LiteLLM that quarantines compromised package backdoors using bash-based system tools without Python dependencies.
3/27/2026
Vibe physics: The AI grad student
Harvard physics professor reports supervising Claude through real theoretical physics research calculations without manual file manipulation.
3/27/2026
Show HN: Pentesting Tool Using Claude
Claude Bug Bounty is an AI agent harness using Claude for autonomous pentesting that reasons about tests, validates findings, and generates bug bounty reports.
3/27/2026
Show HN: I got tired of biased (all) news so I made Claude cosplay each side
Multi-agent system using Claude to analyze news from different perspectives, synthesizing reports with confidence tags across 7 analytical lenses.
3/27/2026
AI tooling for software engineers in 2026
Survey of nearly 1000 software engineers on AI tool adoption, usage patterns, and effectiveness in 2026.
3/27/2026
Use of Large Language Models in Standards Work
W3C Group Note examining considerations for using LLMs in standards development work, discussing useful applications and potential problems.
3/27/2026
Show HN: Run, build, and govern AI agent workflows for DevOps
Retroshift tool for running and governing AI agent workflows in DevOps with action logging and approval gates.
3/27/2026
Tool to reduce AI tokens and generate more detailed and accurate PRDs
Tool generating detailed PRDs and architecture reviews for multi-repository backend systems using AI and dependency graphs.
3/27/2026
Show HN: "Ask AI" for Your Website
Peam: semantic AI search and chat integration for websites with vector-based retrieval and content ranking.
3/27/2026
Two env vars that fix PyTorch memory creep forever – no code changes
Two glibc environment variable fixes eliminate memory creep when repeatedly loading/unloading large ML models in PyTorch on Linux without code changes.
3/27/2026
Gumbel Max trick for LLM sampling
Explanation of Gumbel Max trick for efficient LLM sampling that avoids computing softmax vectors by adding noise and taking argmax.
3/27/2026
Show HN: Superfast – Cognitive Memory Graphs for Enterprise AI Agents
Superfast: enterprise AI agent framework with FastMemory—Rust engine mapping text into functional ontology for improved RAG at scale.
3/27/2026
Goedel-Code-Prover: Hierarchical Proof Search for Open SotA Code Verification
Goedel-Code-Prover uses hierarchical proof search in Lean 4 for automated formal verification of code, constructing machine-checkable proofs instead of relying on testing.
3/27/2026
Environment Maps: Structured Environmental Representations for Long-Horizon Agents
Environment Maps: persistent representations for long-horizon LLM agents to reduce cascading errors and hallucinations in complex workflows.
3/27/2026
AI-Supervisor: Autonomous AI Research Supervision via a Persistent Research World Model
Multi-agent framework with persistent world model for autonomous AI research supervision enabling agents to verify and refine findings.
3/27/2026
Probabilistic Geometric Alignment via Bayesian Latent Transport for Domain-Adaptive Foundation Models
Bayesian latent transport framework for domain-adaptive foundation models addressing distribution mismatch and uncertainty propagation.
3/27/2026
Enhancing Efficiency and Performance in Deepfake Audio Detection through Neuron-level Dropin & Neuroplasticity Mechanisms
Improves deepfake audio detection efficiency using neuron-level dropout and neuroplasticity mechanisms in deep learning models.
3/27/2026
mSFT: Addressing Dataset Mixtures Overfitting Heterogeneously in Multi-task SFT
mSFT algorithm addresses overfitting in multi-task language model fine-tuning by dynamically adjusting compute budget across heterogeneous datasets.