news

Faster coding shifts the bottleneck to the outer loop

Luca Rossi makes a point about AI coding that matches my thinking: most AI coding workflows optimize the INNER loop, the rapid cycle of describing a feature and generating code. But that's a fraction of software delivery. The OUTER loop (deciding what to build, shaping the solution, releasing, learn

Use code for control flow, agents for reasoning

Give deterministic code everything it can handle (routing, scheduling, data transformation, file writes), and reserve the Large Language Models (LLMs)) for the parts that need judgment (interpreting ambiguous, unstructured input).

Orchestrator tax

According to Rahul Garg (Thoughtworks, on martinfowler.com), the real value of a subagent is what it keeps OUT of the orchestrator's context, not how fast it runs. Context isolation is more valuable than parallelism.

Google Agent Development Kit (ADK)

Google's framework for building AI agents, available in Python (since March 2026) and Go. Version 2.0 is a response to a production reality: agents that orchestrate everything through the LLM prototype well and then fall apart in production, with infinite loops, hallucinated routing, and runaway cos

Nano Banana 2 Lite

Google DeepMind's distilled version of Nano Banana 2, published as Gemini Image Flash Lite. The trade is explicit: give up some quality on highly nuanced prompts, get a massive latency drop; roughly 2-5 seconds per image where base NB2 takes ~30. Pricing lands at about half the full model ($30/M out

APM (AI)

APM (Agent Package Manager) is Microsoft's open-source dependency manager for AI agent configurations. You declare your agentic dependencies (skills, prompts, plugins, MCP servers, agent primitives) in a single `apm.yml` manifest, and `apm install` reproduces the exact same agent setup on any machin

Claude Desktop on Linux

Since late June 2026, the Claude desktop app runs on Linux in beta. All three tabs are there: Chat, Claude Cowork, and Claude Code with parallel sessions, visual diff review, integrated terminal and editor, and live app preview. Same experience as macOS and Windows, on all paid plans (claude.ai subs

Context size reduction

The context-window arms race runs in one direction in the marketing (bigger!), but vendors sometimes quietly go the other way. OpenAI reduced the Codex model's context size from 372k to 272k tokens, and because the Codex CLI is open source, the change surfaced as a GitHub pull request rather than an

Treat AI context as a budget to manage

The Context Window is a BUDGET, not a capacity to fill. Every token in context costs money (you pay for it again on every turn), latency, and attention (quality decays as the window fills; context rot sets in long before the hard limit). Filling the window because it's there is spending a budget bec

Ori Eval

Ori Eval is OpenRouter's tool for writing your first eval. Their pitch, and it's the right one:

HuggingNews

HuggingNews is a news wire for AI and open source, where the pipeline itself writes the stories. Its own summary: "Headlines and news articles, written from primary sources." The site framing is simpler still: what is happening in AI today?

OpenWorker

OpenWorker is an open-source "AI coworker" announced by Andrew Ng, built with Rohit Prasad. It's an agent that doesn't just chat with you, but delivers FINISHED work. Polished documents and reports, Slack messages, calendar updates, customer briefs, alert triage; tasks completed across your files an