2026-02-27: Giving Agents Hands, and Systems Over Memory

When you strip it down, today was really just two things:

1. Tightening up my agent toolbox a bit more (Agent Reach + WeChat scraping).
2. Pulling skill management back from “thrown together enough to work” into a maintainable long-term system.

Both seem like dry, tedious engineering chores, but they’re really answering the exact same question: What role do I actually want AI to play in my life and work?

1. Giving agents real reach

Got Agent Reach installed, with 7/12 channels working: GitHub, YouTube, Twitter, Exa search, web fetching… These aren’t gimmicks. When I’m writing something or making a call, they determine whether the agent is just good at chatting, or whether it can actually go pull evidence.

I’ve grown increasingly tired of output built purely on raw generation: the language is pretty, but completely ungrounded. Being able to read web pages, look up references, and pull information from GitHub—that’s what makes it feel like an actual external brain.

2. WeChat scraping: from “it runs” to “it’s usable”

Scraping WeChat articles is essentially about carving out a reliable content ingestion channel for myself.

I ended up building two versions:

• Local Python (BeautifulSoup + custom UA, delays, basic anti-scraping evasion)
• Cloudflare Worker (routing requests through the global edge with rotating IPs, without exposing my local machine)

The most interesting part wasn’t the technical stack, though. It was the moment I realized that with writing and input, the hardest part isn’t expression—it’s consistently securing high-quality raw material.

If my inputs are fed entirely by algorithms over the long run, I’ll turn more and more into the kind of person algorithms love: shallow, hurried, emotional, and impatient.

I don’t want that.

3. Skill management: systems over memory

The real win tonight was standardization:

• Hardcoded the rule in AGENTS.md: all skills must be managed through skill-library—no rogue skills sprouting up in random workspaces.
• Wrote check scripts, daily audit scripts, and a migration queue.
• More importantly: turned multi-agent syncing into an explicit architecture (SSOT + adapter pattern), instead of relying on my brain to remember that “Claude needs a copy, OpenClaw needs a symlink, and Codex needs indexing.”

I used to push back hard against this kind of “institution building”—it felt too corporate, where red tape crushes everyone.

Now I look at it differently: Process isn’t about policing people; it’s about keeping the system running properly even when I’m exhausted or having an off day.

That’s especially vital when you work alone. An individual’s energy naturally fluctuates, but a well-designed system absorbs the swings.

4. In the end: not faster, just steadier

Wrapping this up today made me even more sure of one thing:

I’m not chasing “faster output.” I want “steadier accumulation.”

Faster feels great in the moment, but leaves you empty afterward. Steady is slow, but one day you look back and realize you’re standing on completely new ground.

That’s it.