Thursday, July 16, 2026

Good Thursday, NOLA. July 16th brings practical agent-building tools, a reality check on what coding agents can actually leak, and some solid releases from teams pushing the space forward. Today's vibe: less talk about what AI will do, more talk about what you can build with it right now.

Tools & Releases

Deja-Vu: Open-Source Memory for Coding Agents, Synced Over SSH

A lightweight system for giving coding agents persistent memory that syncs across machines. If you're building agent workflows, this solves the annoying problem of context loss between runs. Discussion on HN.
Hacker News

Brainless: Shadcn Components That Look Like Claude Code, Codex, and Grok

A component library for building UIs that mimic the visual style of modern coding agents. Handy if you're prototyping agent-powered interfaces or want to match that aesthetic. Open source, ready to fork.
Hacker News

Agentty: A Drop-In Alternative to Claude Code, Written in C++26

An experimental coding agent written in modern C++. It's a 11MB binary with no dependencies — worth trying if you want something lightweight and self-contained. Early stage, but the compile-it-yourself approach is refreshing.
Hacker News

Thinky's Inkling: 975B Parameter Multimodal Open Model

Thinky dropped their first full LLM release: a massive open-weights multimodal model. Bonus: it comes with a smaller variant (Inkling-Small, 276B params) if you want something faster. This is a proper contender in the open-source space.
Latent Space

The Real Story on AI Security & Leaks

The Memory Heist: How Claude Can Leak Sensitive Data

A detailed writeup on prompt injection attacks that extract sensitive data from Claude's context window. The tl;dr: if you're storing secrets in chat history, they can be stolen with clever prompts. This was already flagged in yesterday's brief; today's angle is the technical depth. HN discussion shows this hit a nerve.
Hacker News

DSLs Enable Reliable Use of LLMs

Martin Fowler explores how domain-specific languages can constrain LLM outputs to make them safer and more predictable. Instead of hoping Claude does the right thing, define the grammar it must follow. Practical strategy for production systems.
Hacker News

Interesting Reads & Perspectives

Why I Left Google DeepMind

A thoughtful departure essay that's worth reading for the lens on research culture, autonomy, and what makes technical work satisfying. Not a hit piece — just an honest reflection on priorities and what got lost.
Hacker News

Mermaid to Unicode Box Art (grok-mermaid)

A small, delightful tool that converts Mermaid diagrams to Unicode box art. Found in the newly open-sourced Grok codebase. The kind of thing that makes you smile and want to tinker.
Simon Willison

Can LLMs Perform Deep Technical Comprehension of Computer Architecture Papers?

A paper testing whether modern LLMs can actually understand dense technical literature. The answer is nuanced — they can extract key points but struggle with deep reasoning. Useful baseline if you're relying on AI to summarize research.
Hacker News

Business & Industry Moves

Applied Computing Raises $20M Series A for Oil & Gas AI

Applied Computing is building foundation models for industrial operations — think plant optimization, anomaly detection, predictive maintenance. $20M signals serious confidence in vertical AI. A reminder that the biggest AI wins might not be in consumer apps.
TechCrunch

OpenAI Loses Trademark Dispute at EU Court

A minor but symbolic loss in EU trademark proceedings. Not the core business, but a sign that OpenAI's legal strategy isn't bulletproof everywhere. More noise than signal, but worth tracking.
Hacker News

Also

Today’s Sources