ACW #10: GLM-4.7-Flash, Cursor 2.4, and Claude Code Tasks

Welcome back to Agentic Coding Weekly. Here are the updates on agentic coding tools, models, and workflows worth your attention for the week of Jan 18-24, 2026.

Z.ai released a smaller version of their SOTA model. It's a 30B-A3B MoE model and based on benchmarks, the strongest in its class. On SWE-bench Verified, it scores 59.2%. For comparison, Qwen3-30B-A3B-Thinking-2507 scores 22.0% and GPT-OSS-20B scores 34.0%.

If you're using open-weights models for coding, this would be a good replacement for smaller tasks like title generation. Similar to the role played by Haiku models in a Claude Code session. Support has been added to llama.cpp and ollama.

Check the model details on Hugging Face.

Sweep AI released a small 1.5B parameter model for next-edit autocomplete. Runs locally and works with their JetBrains plugin. In the technical blog post, I found the training data preparation strategy and training pipeline sections interesting and worth reading if you're interested in how these models get built.

Todos have been upgraded to Tasks, inspired from projects like Beads.

Tasks are stored in the global claude folder ~/.claude/tasks/<list-id>/ in json format. To make tasks persist across sessions, you need to use the CLAUDE_CODE_TASK_LIST_ID env var. Read the announcement.

So for example, before doing a planning session for a feature, instead of launching claude, run:

CLAUDE_CODE_TASK_LIST_ID=feature-xyz claude

and the tasks will be created inside ~/.claude/tasks/feature-xyz directory. Then you can sequentially work on tasks one-by-one in fresh sessions. Or you can parallelize implementation by working on different tasks which are not blocked by any other tasks in different sessions.

Slash Commands has been merged into Skills. Going forward, when thinking of making a slash command, make a skill instead. Makes sense cause explicitly invoking any skill is basically like running a custom prompt. Codex already merged these two. Read the announcement.

The VS Code extension is now GA. Check the guide to Claude Code in VS Code.

Cursor added subagents, agent skills, and image generation features in 2.4 release. Also added a Q&A tool which agents can use to ask clarifying questions in any conversation. Check the 2.4 changelog.

When building websites with LLMs using coding agents like Claude Code, they often have the same violet-purple color or gradient, and the UI looks like something immediately recognizable as AI-generated. They generally converge towards using common fonts and colors, making all the UI look kind of same.

Anthropic has an official frontend design plugin which you can install. It gets installed as a frontend design skill and helps you generate better UI. It's not a full-fledged replacement for having your own design system, but when you're building small tools for yourself or quick single-page websites, this helps create better UI.

To install, run /plugin inside Claude Code, select frontend-design, and then press Enter to get the install options.

Here's an example. I wanted to build a simple single-page tool that takes a URL as input and shows all the open graph properties like title, description, image, etc.

The first image below (on the left) shows the HTML page generated by Claude Code with Opus 4.5 without any frontend design skill. The second image (on the right) shows the result using the exact same prompt but with the frontend design skill enabled.

Anthropic released best practices for Claude Code last week. Covers almost every stage of working with Claude Code. I'd recommend checking at least the first section: "Give Claude a way to verify its work". Read the post.

This advice to avoid a common pattern is something I learned early. I always try getting things right in one-shot:

OpenAI published a guide on how Codex CLI's agent loop works. Covers building the initial prompt, multi-turn agent loop, and performance considerations like prompt caching and their special compaction endpoint. Read the post and the HN discussion.

OP sees a big gap between the hype online and what he's able to achieve. Asks for real evidence, beyond hype, that agentic coding produces net-positive results. Better to read all the opinions in this thread yourself. Check HN comments.

People share their tips like codify everything, embrace Claude Code for web, use AI as pair reviewer, and more. Read HN discussion.

Gas Town is an experimental and expensive take on IDE from Steve Yegge. It lets you coordinate multiple Claude Code agents working on different tasks. There are lots of moving parts and it's hard to understand how everything fits together. This blog post helps with that.

Gas Town isn't going to be useful for most of us but the patterns used in this project, I believe, are something that's going to show up in dev tools in the future. Worth keeping an eye out and understanding the basic design patterns. Read the post.