MolmoWeb is Ai2’s open-source browser agent — 8B parameters, Apache 2.0, no API key required. It scores 78.2% on WebVoyager and beats GPT-4o-based agents on multiple benchmarks. Here’s how to get it running locally. System requirements: GPU with at least 16GB VRAM (for 8B model) or 8GB VRAM (for 4B model) Ubuntu 20.04+ or macOS 12+ (Linux recommended for GPU support) Python 3.10+ Chrome or Chromium browser installed Step 1: Clone the Repository git clone https://github.com/allenai/molmoweb.git cd molmoweb Step 2: Create a Virtual Environment and Install Dependencies python3 -m venv molmoweb-env source molmoweb-env/bin/activate pip install -r requirements.txt The requirements include PyTorch, the Transformers library, Playwright for browser control, and Pillow for screenshot processing. The full install typically takes 3–5 minutes on a good connection. ...

Claude Code Ultraplan offloads your most complex planning tasks to a cloud Opus 4.6 session for up to 30 minutes — while you keep working locally. Here’s how to use it in five steps. Prerequisite: Ultraplan requires the $400/month Claude Code plan tier. Step 1: Open Ultraplan from the Command Palette In Claude Code, open the command palette (Cmd/Ctrl + Shift + P) and type Ultraplan. Select “Claude Code: Start Ultraplan Session”. ...

The AI agent framework landscape has shifted significantly in the last 90 days. The framework you chose for your agentic project in January 2026 might not be the right choice in April — and one major player has effectively retired. This is the freshest comparison available as of this morning, covering 8 SDKs and 3 inter-agent protocols. The 8 Frameworks 1. Claude Agent SDK (Anthropic) Anthropic’s first-party SDK for building agents on Claude. Deep integration with Claude Code’s subagent architecture, native MCP tool support, and CLAUDE.md context injection. Best choice if you’re building Claude-native workflows and want tight model-SDK alignment. Not model-agnostic. ...

Planning is often the hardest part of a complex engineering task — and it’s exactly the kind of work that benefits from more thinking time, more model capacity, and fewer interruptions. Claude Code’s new Ultraplan feature addresses all three. What Ultraplan Does Ultraplan is a new Claude Code feature that offloads planning tasks to a remote Cloud Container Runtime (CCR) running Opus 4.6 in plan mode for up to 30 minutes. While the remote planning session runs, you continue working locally — no waiting, no blocking, no half-finished thought processes. ...

If you’ve been building Claude Code subagents, you’ve probably hit the same friction point: you spawn an agent, and then you have to manually kick it off with a first message. As of the latest Claude Code release (v2.1.9x, published this morning), that extra step is gone. What’s New: initialPrompt in Frontmatter Claude Code now supports a new initialPrompt field in YAML frontmatter. When an agent is invoked as a subagent, this prompt is automatically submitted as the first turn — no human input required. ...

OpenAI’s Codex just got a major upgrade at the model level. As of April 4, GPT-5-Codex is the default model across Codex CLI, the Codex IDE extension, and Codex cloud environments. This isn’t GPT-5 — it’s a distinct variant, purpose-built for agentic coding workflows. What Is GPT-5-Codex? GPT-5-Codex is a GPT-5 variant optimized specifically for the demands of autonomous coding agents. Where GPT-5 is a general-purpose model, GPT-5-Codex is trained and tuned for: ...

Claude Code’s latest release (v2.1.9x) adds initialPrompt — a YAML frontmatter field that auto-submits a first turn when a subagent is spawned. This guide shows you exactly how to use it. What You’ll Accomplish By the end of this guide, you’ll have subagents that start executing immediately when invoked — no manual prompting required. Prerequisites Claude Code v2.1.9x or later (claude --version to check) An existing Claude Code project with at least one subagent definition Basic familiarity with YAML frontmatter in agent files Step 1: Update Claude Code First, make sure you’re on the latest version: ...

The Allen Institute for AI (Ai2) just dropped something the open-source AI community has been waiting for: a fully open, genuinely capable web browser agent that can go head-to-head with GPT-4o-based systems — at 8 billion parameters. It’s called MolmoWeb, and it’s available right now on Hugging Face under Apache 2.0. What MolmoWeb Actually Does MolmoWeb is a multimodal web agent. You give it a natural-language instruction, and it autonomously controls a real web browser: clicking, typing, scrolling, navigating, filling forms. It understands the web visually — through screenshots — rather than through structured DOM parsing. ...

When Anthropic announced on April 4 that Claude Pro and Max subscribers could no longer use their flat-rate plans through third-party agent frameworks, the developer community felt it immediately. More than 135,000 OpenClaw instances were running when the restriction took effect. Now the creator of OpenClaw — who joined OpenAI in February — has gone public with a sharp response, calling the decision a “betrayal of open-source developers.” What Anthropic Actually Did The policy change is straightforward: Claude’s subscription tiers were designed around conversational use. A human opens a chat window, types a query, reads a response. Agentic frameworks operate differently — a single OpenClaw instance running autonomously through a day can consume the equivalent of $1,000 to $5,000 in API costs. ...

Anthropic’s accidental Claude Code source leak, first reported last week, has had a consequence that security researchers were quietly warning about: someone used the exposed code to find a real, critical vulnerability. This is distinct from the Vidar malware campaign that exploited brand confusion around the leak (also covered here previously). That was opportunistic social engineering — attackers leveraging the story of the leak to distribute malware. What SecurityWeek is reporting now is different: researchers with access to Claude Code’s 600,000-line codebase — exposed via npm source maps — used that access to conduct legitimate offensive security research and found a critical functional vulnerability. ...