Good morning, and welcome to today’s Techzip.

The biggest shift is that AI agents are moving beyond one-off prompts: OpenAI is testing a persistent Codex, Thinking Machines is training models with task expertise, and Anthropic is pushing agents toward physical equipment.

Add a warning from 100+ organizations about AI-enabled cyberattacks, and the message is clear: capability is expanding fast, but memory, access, and control are becoming the real product questions.

WIRED reports that OpenAI is testing a Persistent mode in the Codex CLI. The agent is designed to keep working until it is “put to sleep,” reuse prior interactions and knowledge, create follow-up tasks, and message the user selectively. OpenAI has confirmed the testing, but has not announced a launch date.

The shift is not simply that an agent can run for longer. It turns an agent from a tool that completes a prompt into an ongoing operating layer: one that remembers context across sessions, notices unfinished work, and returns to it. But persistence is not autonomy. The agent does not gain new authority on its own, and actions outside its allowed scope still require approval.

That makes the sleep button, memory boundary, and approval flow part of the product lifecycle, not minor settings. A persistent agent becomes genuinely useful only when a team can answer four questions: What may it remember? What may it follow up on? When must it ask? And who can put it to sleep?

Thinking Machines Lab audited 2,500 examples from BIRD and says 61.1% contained at least one error, spanning the question, the SQL answer, or the supporting knowledge. Rather than simply giving a model more tools and asking it to work things out, the team had experts review the data, define what a correct answer looks like, and use those standards to further train Kimi-K2.6.

The lab reports 91.37% under its default run and 92.97% when sampling multiple answers and selecting the best result. Those are company-reported results, not proof that Text-to-SQL is solved. Still, they point to something useful: for a narrow enough task, teaching an AI the standards of strong practitioners can work better than making it infer those standards while doing the job.

The practical point is that expertise does not live only in prompts or final review. If a team knows which mistakes recur and what a correct result looks like, it can put those standards into training data and help the system get more right from the start.

More than 100 organizations, including OpenAI and Anthropic, signed a warning that companies and critical infrastructure may have only months to prepare for AI-enabled cyberattacks. It is a collective warning, not evidence of a confirmed attack timeline.

But the concern is not limited to novel vulnerabilities: excessive permissions, misconfigurations, unpatched software, weak authentication, and legacy systems can all become cheaper entry points when attackers have AI help.

The story is therefore not only about defending against AI. It is also about using AI for defense before attackers scale. The signatories call this the “defender’s window”: companies raise their security bar, security vendors continuously test defenses and share validated threat intelligence and playbooks, and governments coordinate support for under-resourced infrastructure.

Finding the right payment processor shouldn’t take weeks of wrong vendors and inbox spam. There’s a faster way with Top Provider. Answer a few questions about your business: your industry, your volume, and how you take payments. In minutes, we match you with processors that fit your specific situation. No research rabbit holes. No chasing quotes. The right fit, faster than you thought possible. Get started with Top Provider today.

Compare Payment Processors for Your Business with Top Provider

Today, I bring to you this guide from SpaceXAI designer - John Bai.

He describes using Grok Bot as a team of AI design agents to design Grok Bot. The point is not to replace the designer. It is to prototype faster, explore more directions, and reduce repetitive production work while the human still decides which version is actually good enough to keep.

The design loop has four parts:

The loop is: idea → agent prototype → designer response → agent revision → designer judgment → repeat. The real value is not only speed. It is making the cost of trying an idea low enough to explore more directions before committing to one.

A Tsinghua study covering 1,338 post-training runs finds that agents can train, debug, evaluate, and iterate, but rarely reconsider strategy. It is a useful counterweight to the idea that simply giving an agent more tokens will make it find its own way.

This essay argues that the AI debate should focus on nonproliferation and control, not only data centers. It reframes the compute race around a harder question: who gets access to capability, and which systems can keep it bounded?

This conversation looks at the gap between the speed of capital and capability and the investment going into safety. It is a useful long-horizon lens when every day brings another capex number.

This paper argues that policy should target system-level controls rather than model categories alone. It offers a compact framework for reading today’s cyber warning: the risk may begin with a model, but the damage usually happens in the systems connected to it.