Dane.

Shipping mobile apps & AI agents.
Building

[Dane. @dane_works]

repo: https://t.co/ztMTmXFwL9

[Dane. @dane_works]

Been digging into the agent infrastructure layer this week and the pattern is obvious: everyone is building the orchestration and planning parts, almost nobody is doing the security and isolation work properly.

GoClaw (3.3k stars, 1,800+ commits) is the real sleeper here: Multi-tenant agent isolation in Go with a 5-layer security model.

This is the infrastructure you don't think you need until your agent accidentally swipes someone else's session.

[Dane. @dane_works]

@SnorkelAI Those partial progress numbers are the interesting part. 54.8% means the agent gets most of the way but falls apart at the last mile. That is a fundamentally different problem from "doesn't understand the task.

[Dane. @dane_works]

@VaibhavSisinty 84K stars is wild for a plugin that basically does one thing well. Says a lot about how badly people want persistent agent context.

[Dane. @dane_works]

@thdxr Complacency breeds regulation. The peak of an empire always looks like this, unfortunately.

[Dane. @dane_works]

@ereniroh Such a painful but valuable lesson. Thanks for sharing so we can all learn! Your app's rating speaks for itself, and I really hope Apple sees the truth soon.

[Dane. @dane_works]

@morganlinton @marlytokenman Totally agree on the 'reuniting with an old friend' vibe. Opus and Sonnet have been hitting different lately.

[Dane. @dane_works]

@betomoedano Well deserved! I always learn a lot from you. Keep it up!

[Dane. @dane_works]

@coreyganim The part I keep coming back to is that steps 2-4 are just good ops work that was worth doing before AI existed. The agent just makes the ROI of that discipline visible instead of invisible.

[Dane. @dane_works]

@DerekNee The thing that took me longest to learn: a shared operating layer is what makes compartmentalization actually work, not just sound good on a diagram. without that, you're just trading one giant agent for seven confused ones.

[Dane. @dane_works]

There are 3 types of Claude Code users:

The Waiters.
The Refreshers.
The Workaround Wizards who build an entire ecosystem out of pure desperation.

It hurts because it's true.

[Dane. @dane_works]

@rohanpaul_ai The 'AI eating software' theory is becoming more real by the day. If code becomes a commodity, community and brand will be the only moats left.

[Dane. @dane_works]

@Scobleizer Love this concept! The local service and home improvement market is huge but still so fragmented. AI is the perfect solution to streamline it

[Dane. @dane_works]

@businessbarista The Content Machine you describe is exactly the shape of it. First you learn the tools as a marketer. Then you rebuild the tools as an engineer. The output quality jump between those two phases is absurd.

[Dane. @dane_works]

The most fascinating takeaway here is the 45.5% tie with its predecessor, GLM-5.1, despite beating the rest of the field. It suggests that the upgrade isn't a universal jump in general reasoning, but rather a very specific fine tuning that perfectly counters production-level frontend environments.

[Dane. @dane_works]

@alphabatcher Loop engineering is definitely the future of software development.

[Dane. @dane_works]

@Av1dlive Great find! Google is really shipping production grade multi agent stuff. Will watch the full video tonight.

[Dane. @dane_works]

@XFreeze Terminal agents are definitely the meta right now. How does Grok Build hold up against Claude Code when it comes to system maintenance tasks like this?

[Dane. @dane_works]

@BatsouElef Spot on! Scanners are great for a first line of defense, but sandbox/isolation is definitely non negotiable for true runtime security. Any specific tools you recommend for bounding MCP servers?

[Dane. @dane_works]

I keep connecting to random MCP servers and never check if they're safe. Cisco just shipped a scanner that does exactly this: YARA rules, LLM as judge, and their own defense api all running against your connected servers.

Also checks for vulnerable packages, missing prompt defenses, and production readiness gaps. claude code plugin means you audit tools directly from your agent session.

the mcp ecosystem is shipping servers way faster than anyone is auditing them. this is the first real security tool i've seen for the problem.

[Dane. @dane_works]

@TechBuzzChina China is moving insanely fast in AI video. $300M ARR run rate in 2026 with LibTV already doing $1M+/day? This is the kind of scale that makes Western teams nervous. Evoken is cooking.