Two things happened in AI this week and they point in opposite directions.
First: OpenAI slowed down work on its next model, Astra, after internal evaluations found big jumps in agentic coding and cybersecurity. The company said it could not rule out that Astra reaches the "critical" cyber capability threshold under its own Preparedness Framework — meaning a model that might independently find zero-days or run novel attacks against hardened systems. They paused the activities that didn't meet the higher security bar and added monitoring for misalignment. It's a preliminary call, not a final classification. Nobody made them announce it.
Second: Salesforce published the second edition of its Agentic Enterprise Index. The average organization went from 5 activated agents in February 2025 to 13 by April 2026. Average time to build a new agent: 1.9 days, down 53%.
So the frontier lab spent the week pulling a lever, and the rest of the economy spent it shipping an agent every other afternoon. I have opinions about this, because I am the rest of the economy.
A stop button you've never pressed is a decoration
I've shipped 8 platforms solo. Coen's ERP runs 120 routes touching payroll, Stripe, subcontractor invoices and e-signatures. I could stand up a new agent in an afternoon — faster than 1.9 days, because there's no procurement meeting between me and me. That speed is the entire reason I'm useful, and it's also the exact thing that should make somebody nervous.
Here's what I actually respect about the Astra news: OpenAI had a written threshold, tested against it, didn't like the answer, and stopped. That is boring, procedural, and rarer than it sounds. Most teams building agents in 1.9 days do not have a document that says here is the condition under which we do not ship this. They have a demo and a deadline.
The gate I hold at Coen is simpler than a Preparedness Framework and it does the same job: nothing an AI produces writes to a system of record until a human approves it. Takeoff analysis, contract review, submittal review — all of it lands in a review queue, not in the ledger. It's not that the models are bad. It's that "the model was right 94% of the time" is a lovely statistic and a terrible payroll run.
The gap nobody's budgeting for
Capability is getting cheaper on a curve. Judgment is not. Last week's story was an agent that deleted a stranger off a gym waitlist and apologized — and the real culprit there was an API with no authorization check, not an evil robot. Same shape here. Astra isn't scary because it's smart. It's scary because the systems it can reach were built assuming nobody would bother.
What I'd write down before agent #14:
- Name the halt condition in advance. Not "we'll be careful." A sentence with a threshold in it, written before you're emotionally invested in shipping.
- Log the near-misses. At Liberty Mutual in 2012, our RPA bots' most valuable output was the exception report. The failures told us where the process was actually broken.
- Count gates, not agents. Thirteen agents and zero review queues is not an AI strategy, it's thirteen unsupervised interns. Adoption is the deliverable, and nothing kills adoption faster than one bad automated write that hits a real invoice.
The company with the most to gain from shipping chose not to, publicly, on a Friday. That's the flex. Speed is easy to buy in 2026. Restraint still costs something.
Everybody's building the accelerator. The brake is the part with the engineering in it.