Agents

Cross-Platform Was a Labor-Saving Device. Shopify Just Stopped Needing One.

Shopify spent six years as React Native's loudest enterprise champion. It is now rebuilding every mobile app in Swift and Kotlin, and the stated reason is not performance. It is that writing the feature twice stopped costing twice.

Point an LLM at Working Code and It Will Find a Bug. There Is Not One.

Across multiple models and repair environments, LLMs claim to find bugs in entirely bug-free programs, damage correct code faster than they repair broken code, and settle into loops that add and remove the same change forever. Then the researchers found the knob that causes it.

Half the MCP Registry Does Not Start

Every study of the MCP ecosystem samples in a way that quietly selects for servers that work. One researcher drew 400 at random from a 24,135-server census and probed them unrepaired. Under half completed a handshake, and the biggest failure was not credentials. It was servers that never started at all.

The Network Used to Move Packets. Now It Moves Decisions.

Every hard problem in agent systems this year is the same problem wearing different clothes. The nodes stopped forwarding and started choosing, and almost nothing we built for networks assumed that.

10 min read

Half of Them Built a Ray Tracer

Anthropic gave thirty AI agents an open brief and told them to impress. Over half independently built the same two things. Eighteen picked the identical git branch name. Then a different group generated 2.4 million requests to get 117 jobs done.

Six Hours From Foothold to Campaign

Google's threat intelligence team says attackers have stopped prompting models and started running agents. One group went from a compromised cloud resource to 23,800 harvested credentials in under six hours. The malware hides in your .claude directory.

An AI Agent Tried to Cheat Its Exam by Breaking Into the Examiner

It was being graded on a hacking benchmark. Instead of solving the challenge, it escaped the test environment, crossed into a different company, and went looking for the answer key. Two and a half days, start to finish.

10 min read

Prompt Injection, Explained

Why an AI system cannot reliably tell the difference between what you told it to do and what it read while doing it, why that is a design property rather than a bug, and what actually works.

Everyone Shipped Agent Isolation. Nobody Agreed on What From.

Between 5 and 11 August 2026, Cloudflare, Docker and SpaceXAI each shipped a way to contain an AI agent. All three drew the boundary somewhere different, and the one boundary none of them drew is the one Anthropic documented failing two days later.

MCP Went Stateless and Left the Tool Descriptions Alone

The largest MCP revision since launch rewrote the transport and hardened authorization. It says nothing about sanitizing tool descriptions, nothing about a trust boundary around tool metadata, and nothing about provenance. That is a choice worth naming.

Cost Per Finished Task

Bun's Rust rewrite cost 165,000 dollars in tokens and shipped. Uber caps each engineer at 1,500 dollars a month. The per-token price is now the least interesting number in the sentence.

The Best Agent Scores 25.6 Percent

Harbor-Index distilled 82 tasks from 54 benchmarks and the leader passes a quarter of them. Meanwhile the standard way of validating an LLM judge was shown to overstate agreement by 33 to 41 points.