Axentia

genAi dev studio

[Axentia @axentiatech]

🚀 We're hiring at Axentia!

🔹 Founding Engineer: Build cutting-edge AI products with the founder.

Tech: Next.js, TypeScript, Prisma, AI/LLMs (OpenAI, Anthropic, Vercel AI SDK).

💼 Perks: Flat hierarchy, revenue share, direct impact.

🌐 Apply now: https://t.co/09VgUKvIrq

[Axentia @axentiatech]

Plugin4Shell: vibe coders using Claude Code, Copilot, or Gemini CLI need to be careful.

Zero clicks. No approval prompt. A malicious plugin update runs code on your machine while your agent says everything's fine.

Here's what happened 🧵
https://t.co/q96YVsmidg

[Axentia @axentiatech]

Your AI-built app can look finished and still expose customer data, API keys or payment flows.

VibeAudits manually reviews AI-built SaaS apps to find the security issues automated checks miss.

Launching soon? Get it reviewed.

https://t.co/1S89hjeHg4

[Axentia @axentiatech]

Your AI-built app can look finished and still expose customer data, API keys or payment flows.

VibeAudits manually reviews AI-built SaaS apps to find the security issues automated checks miss.

Launching soon? Get it reviewed.

https://t.co/1S89hjeHg4

[Axentia @axentiatech]

I would compare the fitted model with a standard quantization on the same prompts, context and hardware.

Measure task accuracy, latency, memory peaks and failure rate.

Choose it because the results are better, not because the suitcase closes.

https://t.co/Qek7ODVkiq

[Axentia @axentiatech]

A local AI model fits in memory. Good.

That only proves it can load.

It does not tell you whether the answers are accurate, long prompts work or the app stays stable under real use.

I would not choose a local model from file size alone.

[Axentia @axentiatech]

Shoehorn reports more than 99.9% budget use in practice.

That means it can fill the available byte budget closely. It does not mean 99.9% of model quality is preserved.

A smaller file is useful only if the product still does its job.

[Axentia @axentiatech]

I would test one thing before putting an agent into production:

Can the team replay a failed run and explain every important decision?

If not, it is still a demo.

More on DeepSeek Harness:

https://t.co/zvcriXYsy2

[Axentia @axentiatech]

An AI agent edits the wrong file.

You can see the bad change, but not why it happened.

Which prompt did it receive? Which tools could it use? What did the command return?

If your team cannot answer those questions, it cannot operate the agent safely.

[Axentia @axentiatech]

The log creates its own problem.

Tool output can contain source code, customer data, private paths or credentials.

Before keeping session history, you need rules for redaction, access and retention. An audit trail should not become a collection of secrets.

[Axentia @axentiatech]

The goal is not an AI agent that never stops.
It is an agent that knows what it can do, how success is measured and when it must stop.
I wrote more about Headlong, its costs and its limits here:
https://t.co/xzFcdHjpTE

[Axentia @axentiatech]

A company switches on an AI agent on Friday and tells it to keep improving the product over the weekend.
On Monday, the agent is still running, the API bill has grown, and the team has a pile of changes nobody has reviewed.
This is the problem with always-on AI.

[Axentia @axentiatech]

I would not start by running an agent permanently.
I would give it one task, one isolated workspace, a fixed budget and eight hours.
Then I would review what it produced, how much it cost and how often someone had to intervene.
It should prove that persistence is useful.