$0 extension turns your Claude Code terminal into something you've never seen it do before.
You open a normal VS Code terminal. Nothing looks different - until a tiny pixel character spawns inside it.
That's Pixel Agents, a free extension that gives every Claude Code terminal its own miniature office.
The character isn't decoration. It mirrors the agent live.
Writing code, it types. Reading a file, it reads. Stuck waiting on you, a speech bubble pops up over its head.
Open 3, 4, 5 terminals at once and the office fills up. The agents walk around. They coordinate with each other. Visually. On screen. In real time.
Behind the scenes it just reads transcription log files, so none of this touches your actual workflow.
FREE. OPEN SOURCE. 1 install and your terminal stops being a wall of text - it becomes a room you can glance at and instantly know who's doing what.
You stop reading logs. You start watching a tiny office run itself.
One AI agent replaced a content manager and now runs 7 sub-agents alone.
They call it Nolan. Easier to say "ask Nolan" than "ping the automation pipeline."
Founders table gave every department its own AI head like this. Each one manages 5 to 7 sub-agents underneath. You give the head one task. It splits the work, runs it, hands back one clean answer.
Above them: 1 chief agent for executives. It queries every department head at once and merges the replies before you ever see them.
Employees never touch the sub-agents. They open a chatbot, pick from 5 to 7 examples, and ANY NEW HIRE IS RUNNING IT WITHIN A SINGLE AFTERNOON.
YOU DON'T LEARN THE TOOLS. YOU JUST TALK TO THE HEAD.
Voice works too. Call the agent like a person.
Building the full org out over the next 12 months.
You stop managing agents. You start managing names.
Zhipu just cut Claude's coding lead to 4 points with a free 1,000,000-token model.
GLM-5.2. Fully open source. MIT license. No usage caps. No regional lock. No bill.
On Terminal-Bench 2.1: GLM-5.2 scores 81.0. Claude Opus 4.8 scores 85.0. That's the entire gap. 4 points.
On PostTrainBench it already beats GPT-5.
SWE-bench Pro: 62.1, up from 58.4 in the last version.
RUN IT LOCAL. 40 billion active parameters out of 753 billion total. ZERO API cost. ZERO rate limits. ZERO waitlist.
Don't want to self-host? 1 command. 1 API key. Live inside your coding agent in under 2 minutes.
1,000,000 tokens of context. Drop in a full repository. No chunking. No summarizing half your codebase just to make it fit.
2 reasoning modes ship with it, high and max, so speed trades for depth on command.
Silicon Valley spent billions building that 4-point lead.
China just handed it out for free.