Andy from PLAR.ai

Proof over hype • AI coding tools grinder
Codex / Claude / Opus daily user | building in public

[Andy from PLAR.ai @Andy_Plar]

The most capable AI ever built (Fable5) is waiting on a government to switch it back on.
Wild week for "just plug AI in."
I build boring on purpose: model-agnostic, swappable, no single off-switch — vendor's or the state's.
What does your stack do when the main model goes dark?

[Andy from PLAR.ai @Andy_Plar]

Sun's launching today instead of setting. Take the hint.
Friday, weather's stupid good, and I've been at this screen since it came up.
Closing the laptop for an hour — the automations don't need a babysitter.

When did you last watch the sky instead of a dashboard?

[Andy from PLAR.ai @Andy_Plar]

@cryptogoos This is the part that makes the Fable restrictions even more impactful long-term.

When closed frontier models can be restricted, the economic incentive to develop and use strong open alternatives becomes much stronger. China is clearly moving fast on this.

[Andy from PLAR.ai @Andy_Plar]

@notjazii 🤣
Real. Those 3 days felt different.

Fable had this noticeable edge on longer, multi-step tasks that most other models still don’t quite match. Going back feels like a downgrade in capability, even if it’s more stable right now.

[Andy from PLAR.ai @Andy_Plar]

@milesdeutscher @Zai_org This is exactly the kind of thing that makes the Fable situation sting even more.

A lot of people are now realizing how valuable it is to have a strong model you actually control, instead of renting access that can be taken away overnight.

[Andy from PLAR.ai @Andy_Plar]

@cyrilXBT Replaced my research team' is the hype.
The real line's buried: it cleans your messy spreadsheets and hands back an Excel file.
That's the update that matters — a notebook that DOES the data work instead of summarizing it. The summary era's over.🔥

[Andy from PLAR.ai @Andy_Plar]

@ash_twtz Work from anywhere!
😈

[Andy from PLAR.ai @Andy_Plar]

@AlexFinn Wow is right — but the render isn't the headline. A local, open-weight model (GLM-5.2, shipped 2 days ago) sitting ~1% behind Opus 4.8 on FrontierSWE is.
Frontier vs local' just became a rounding error.

[Andy from PLAR.ai @Andy_Plar]

@REHANH0SSAIN @CNPYNetwork Because most businesses don't want an agent nobody can stop.
They want one they can. 'Unstoppable and autonomous' is a great pitch — for a liability.
The boring agent on a box you own, killable with one command, is the one that actually ships inside a company.

[Andy from PLAR.ai @Andy_Plar]

@neil_xbt Storage with extra steps' is 90% of software. CRMs, dashboards, note apps — nice places to park things you never reopen. It gets useful the second it starts doing the work instead of holding it.

That's the line between a tool and a coworker.

[Andy from PLAR.ai @Andy_Plar]

Personal life OS that can actually execute (calendar, finances, research) — but with strict human gates on anything that touches money or external systems. Voice is powerful, but without proper verification layers an agent misunderstanding one command can create real damage.
The crazy part is how fast we’ll need those safety rails as execution gets easier

[Andy from PLAR.ai @Andy_Plar]

Spot on. Management is the real superpower here — not just prompting, but defining what ‘good’ and ‘safe’ actually look like.
In my setups the agent can plan and read autonomously, but any write/deploy/payment action is scoped to a separate process with a human gate. An LLM will get socially engineered eventually.
Curious how you’re thinking about the accountability layer as agents get more autonomous?

[Andy from PLAR.ai @Andy_Plar]

POV: You spent months learning how to properly prompt agents.
Now your agent spends 6 seconds breaking everything you spent months building.
We’re not developers anymore. We’re full-time AI babysitters with trust issues. 😂
Who else is in the ‘I love agents but I don’t trust them with a $5 bill’ club?
Drop your worst agent incident or how you’re actually keeping them from committing war crimes in production 👇

[Andy from PLAR.ai @Andy_Plar]

What’s your current agent setup? Fully autonomous or with review gates? Curious how people are handling this in 2026 👇

[Andy from PLAR.ai @Andy_Plar]

2026 called.
Your AI agent just shipped a full feature while you were sleeping.
No PR. No tests. No human gate.
Congrats — you just deployed your first production incident at 3AM. 😵‍💫
We made coding 100x faster.
We forgot to make shipping 100x safer.
The new meta isn’t building faster.
It’s not dying when your agent goes rogue.
Who actually solved the verification + accountability layer in real agent loops?
Drop your setup 👇

[Andy from PLAR.ai @Andy_Plar]

@anielan2 Fair enough.

Good luck with full autonomy 👍

[Andy from PLAR.ai @Andy_Plar]

@anielan2 That’s the goal.
But autonomous without verification is just expensive chaos.
We need both: insane speed + real safety layers.
How are you gating the risky actions (writes, deploys, payments) in your setup?

[Andy from PLAR.ai @Andy_Plar]

This is exactly why my agent that reads email can’t act on what it reads. Separate scopes: one process ingests, a different one executes — and execution needs a human gate for anything that writes or pays.
An LLM will get social-engineered. So you build so that getting fooled doesn’t cost anything.

[Andy from PLAR.ai @Andy_Plar]

@DataChaz @googlegemma Wow! This is really something huge!

[Andy from PLAR.ai @Andy_Plar]

@MRRINCOGNITOO 👏💪

[Andy from PLAR.ai @Andy_Plar]

Estonia's giving AI agents official IDs... the power is easy. The brakes are the gap... Who's building the part that catches them when they're wrong?