Welcome back to Agentic Coding Weekly. Here are the updates on agentic coding tools, models, and workflows worth your attention for the week of Dec 21 - 27, 2025.
Tooling and Model Updates
GLM-4.7
Z.ai released GLM-4.7, an open-weights MoE model with 355B total parameters and 32B activated. Comes with 200k context length and scores 73.8% on SWE-bench Verified.
Even at 4-bit quantization (Q4_K_M), it'll need about 130GB VRAM to run it locally. Through API, it's $0.6 / $2.2 per million input / output tokens.
Can be used with Claude Code following their guide. Check the GLM-4.7 announcement.
MiniMax M2.1
MiniMax made their SOTA model open-weights. It's a 229B-A10B (229B total parameter with 10B activated) model that scores 74.0% on SWE-bench Verified. This makes it the best performing open-weights model on that benchmark.
Previous leader was DeepSeek-V3.2 at 73.1%. For proprietary models, Gemini 3 Pro sits at 76.2%, GPT-5.1 Codex Max at 77.9%, and Opus 4.5 leads at 80.9%.
API pricing is $0.3 / $1.2 per million input / output tokens. More expensive than DeepSeek V3.2 ($0.28 / $0.42) but much cheaper than GLM-4.7 ($0.6 / $2.2), Gemini 3 Flash ($0.5 / $3), and Claude Sonnet 4.5 ($3 / $15).
Also works with Claude Code. Read the MiniMax M2.1 announcement.
Quick Updates
Claude Code added LSP (Language Server Protocol) tool for code intelligence features like go-to-definition, find references, and hover documentation. Currently not working due to a bug.
Opencode added prompt stashing. It's a quality-of-life update I've always needed. Sometimes I have a half-written prompt I don’t want to lose, but I need to ask something else first. Now I can stash it.
There are 3 commands: "Stash Prompt" (saves current prompt), "Stash Pop" (retrieves most recent stashed prompt), and "Stash List" (shows all stashed prompts).
Anthropic doubled usage limits for Pro and Max plans from 12:00am PST December 25 until 11:59pm PST December 31. Time to vibe code that side project you keep putting off.
Spotlight: 2025 Agentic Coding Reading List
I curated the best community-written posts I found this year on using LLMs and coding agents to write and ship professional code and organized them into following 9 categories (Practical Workflows, Case Studies, Code Quality, and more). 42 posts total.
If you want to catch up on the year quickly, or you’re new to agentic coding and want a solid starting point, this list should give you the main ideas, debates, and failure modes without the noise.
Here are my 11 must-reads:
Claude Code: Best practices for agentic coding (Apr 18): The canonical starting point for Claude Code users.
My AI Skeptic Friends Are All Nuts (Jun 2): Arguments against AI is a fad. Writing style makes it a delightful read.
Software in the era of AI (Jun 19): Andrej Karpathy keynote.
6 weeks of Claude Code (Jul 30): Years of tech debt cleared in a month and a half.
LLMs + Coding Agents = Security Nightmare (Aug 18): Introduces RRT (Refrain Restrict Trap).
Vibe engineering (Oct 7): When seasoned professionals accelerate their work with LLMs while staying accountable for the software they produce.
Two things LLM coding agents are still bad at (Oct 9): No true copy-paste. Can't ask good clarifying questions.
How I use every Claude Code feature (Nov 2): Reference for Claude Code’s entire ecosystem.
You Should Write An Agent (Nov 6): You only think you understand how a bicycle works, until you learn to ride one.
Writing a good Claude.md (Nov 25): Self explanatory.
Your job is to deliver code you have proven to work (Dec 18): Don't shift the burden of verification to reviewers.
Read the entire list of 42 posts:
Subscribe to get weekly updates delivered to your inbox:
Community Picks
Your Guide to Local Coding Models
Explores viability of using local coding models as an alternative to expensive cloud subscriptions. Covers setup process, hardware considerations, and performance trade-offs.
Author changed their conclusion after feedback: "But do I want someone reading this to immediately drop their coding subscription and buy a maxed out MacBook Pro? No, and for that reason I need to correct my hypothesis from ‘Yes, with caveats’ to ‘No’."
Scaling LLMs to Larger Codebases
Argues scaling LLMs to larger codebases requires investing in both guidance and oversight to maximize "one-shotting" and minimize rework.
Guidance: prompt libraries, best practices, codebase maps to provide necessary context. Oversight: elevating human design capabilities and automating verification.
Key quote: Making a prompt library useful requires iteration. Every time the LLM is slightly off target, ask yourself, "What could've been clarified?" Then, add that answer back into the prompt library.
Ask HN: How are you sandboxing coding agents?
Tips and practical insights on running coding agents in sandboxed environments from the HN community. Read HN comments.
That’s it for this week. I’ll be back next Monday with the latest agentic coding updates.