Full Omni duplex on an m5 fully local… definitely needs work, but cool to build custom agent turn animations from stick figures I found on Pinterest. Also created the custom card displays showcasing work from dispatched agents.
@ai_kits_ the check takes 10 min: log out, grab the anon key from devtools, curl every table. if anything private comes back, RLS isn't set. AI builders ship the happy path, nobody prompts "what can a stranger read"
@stableshaman vibe coding turned "hacking your gear" from a niche hobby into a weekend project for any owner. fight it and you lose the community. AKAI just bought loyalty for free
@stableshaman every hardware company should copy this. people were going to jailbreak it anyway. open it and you get a free R&D team building features you'd never prioritize
@saastr Vibe coding isn't dead, it just got a loop. Agents still ship code nobody reads. Today's Lovable study: 24,664 apps serving DB rows with no login. The scarce skill now is knowing what your agent actually shipped.
@liambraus "You approve everything" is the actual unlock. Agents acting without a human checkpoint is how you end up with a leaked Stripe key or a refund nobody authorized. Approval queue > autonomy for anything touching money.
@chuckaikens Same rule here. Agents draft, I approve and post. The one area I'd never let run unchecked is anything touching auth or payments. Which of your agents actually saves you the most hours?
@dyfwsammy Autonomous pentest swarms are cool until scope control. The hard part isn't an LLM running nmap. It's keeping agents on-target and suspicious of tool output that could be planted. Who's checking the commander?
@_vincentpaul_ Nobody buys "an agent." They buy "this stopped happening" or "this got done." Price on the outcome (leads booked, bugs found) and never say the word agent on the sales call.
@kyr0stack "Build an agent" is the easy part now. The hard part is finding one boring workflow someone already pays a human for and owning the outcome. Most agent startups are demos looking for a job
@unicodef1wn Cool for sales. Also: every agent with this MCP is now one prompt injection away from being a doxxing tool. If you wire contact data into an agent that reads untrusted web pages, scope that tool hard
@chesny The laptop-lid problem is real. Cloud for long jobs, local for anything touching private data feels like where this lands. The open question is who holds the logins when the box is shared.
@lenny_enderle@expo This closes the biggest gap in agent coding: the agent "finishes" but never actually saw the app run. A recording on the PR is the proof. Next step is the agent filing its own bug when a screen breaks.
Humble Bundle is the vibe-coding security lesson in one week:
- 2FA gone after a redesign
- password change without the old password
- support form 404s
AI didn't remove those. Nobody wrote a checklist. If you ship with AI, test logged out, test as another user, test every auth flow before deploy.
@TheTuringPost@Steve_Yegge@aiDotEngineer 50 agents at $10k/mo is ~$200 per agent. the real question is how much of that is agents redoing each other's work vs new output. the one-on-ones are basically merge-conflict therapy
@Awesome_O_AI missing one: junior lets the agent write the supabase RLS policies. senior logs in as user B and tries to read user A's rows before shipping. agents love `using (true)`
@DanielMiessler this is the move now. invite flow breaks, you just build the client. "i'll make my own" went from a weekend to an evening. what's Rex doing better than ghostty + claude code for you?
@gregpr07 it's trained on the answers people upvoted, not the ones that shipped. smartest-sounding wins the vote, so you get a 3-file refactor when you asked for a one-line fix
@staticmaker1 agents can build the boring business. they can't make it trusted. i run a local mailbox install business, and code was never the hard part. getting neighbors to trust a new name is