Ergon Insights

09/30/2026

Introducing Tekton by Ergon Insights: an AI engineer inside your business.

Tekton places an Ergon AI engineer with your company to fix and automate the processes that eat your team's week, using the tools you already run. It's remote-first, working with firms anywhere. Every engagement runs under Ergon leadership, and Jason Oglesby owns the result.

Start with a $400 starter project: one time, 1–2 hours, one small process fixed or automated. Then Focused at $2,000 a month keeps your priority workflow moving. Accelerated ($4,000) and Expanded ($8,000) scale up from there.

Book a strategy call:

A remote-first AI engineer embedded in your business, automating the processes that eat your week. Start with a $400 project, then $2K/mo.

09/22/2026

Anthropic published a scorecard for agent oversight last week and then filled in its own row.

Coverage. Review latency. Escalation rate. Three questions any operator should be able to answer about a system they are accountable for.

Worth naming our interest: Ergon sells technology leadership work that includes this kind of discipline, and Anthropic's numbers are self-reported.

What they published for August: about 30,000 agents running at once on their main internal platform, every action passing through a monitor, and 0.002 percent of decisions blocked across more than a billion.

Most companies cannot answer the first question. Not because their monitoring is bad, but because nobody knows how many agent actions exist to begin with. Coverage is an inventory problem wearing a monitoring costume.

The number to remember: roughly 100,000 transcripts flagged in a week, about 50 reaching a person.

That is not a failure. Fifty a week is a job somebody can do. A hundred thousand a week is a wish. Every company adding agents is heading toward that ratio whether they have thought about it or not.

Count what is running before you try to measure it.

09/18/2026

Special thanks to John Hunter for this awesome review of our AI education!

09/11/2026

Never Forget

09/05/2026

On September 1, OpenAI wrote that its new model can "find previously unknown security flaws and develop ways to exploit them across many well-protected systems without a person guiding each step."

That is their sentence. Astra is the first model they have designated Critical for cybersecurity under their own framework.

Then they published a number. Jailbreak refusal went from 59 percent on the previous model to 91.5 percent on this one.

Worth naming our interest: Ergon sells technology leadership work that touches security posture.

Credit first. They published the framework, the threshold and the results before shipping, rather than after a researcher forced the issue. Most vendors do not.

Now invert it. 91.5 percent is a real improvement. It is also 8.5 percent, roughly one attempt in twelve getting past refusal training in their own test.

Refusal is one layer of several. They also describe classifiers, monitoring, tighter limits on high-risk accounts and a round-the-clock red team. They built those because they know a percentage is not a wall.

The mistake is not theirs. It is a buyer reading the headline number as the safety story when the vendor treats it as one control out of six.

Most security programs assume finding a novel vulnerability is slow and expensive. That assumption is what your patch deferrals are priced against. Go look at them.

OpenAI rated Astra Critical for cyber and published the eval: jailbreak refusal 59 to 91.5 percent. Credit for publishing it. Then do the arithmetic they…

09/04/2026

OpenAI released GPT-6 Astra yesterday, and Greg Brockman said "Welcome to the AGI era."

Here is what nobody put in a headline. At least three different things are wearing that name, and the one being written about is not the one on your account.

Access started with a limited set of organizations in OpenAI's Daybreak program. Plus, Pro, Business, Enterprise and API developers get it "in the coming days." The most capable cybersecurity functions are held back for a small group of trusted testers.

Worth naming our interest: Ergon sells implementation work, so a post arguing that announcements are not deliveries sells what we do.

A staged rollout is defensible. If a model can cause harm at scale, releasing it gradually is what a responsible vendor does. But defensible for the vendor and usable for the buyer are different questions.

A roadmap needs a date. A budget needs a price. A security review needs a scope. "In the coming days" gives you none of those.

OpenAI also says Astra set new marks on biology, chemistry, medical and physics evaluations. We went looking for the numbers. Not one report names a benchmark or gives a score. That is a direction, not a measurement.

Plan against what you can call today. Everything else is a rumor with a version number.

OpenAI shipped GPT-6 Astra to Daybreak orgs first, everyone else in the coming days, best functions held back. A roadmap needs a date. Plan against what…

09/04/2026

The owner is no longer the back office.

That's BlindsPlus. Invoicing, the customer portal, payments, and follow-up on blinds.plus. "It saves me over 3 hours a day."

BlindsPlus. Custom window treatments. The operations platform on blinds.plus.

09/03/2026

Nvidia is buying Hugging Face for $12.93 billion. That was announced this morning.

The layer most engineering teams treated as neutral ground now belongs to the company that sells the chips.

Worth naming our interest: Ergon builds model-agnostic architecture, so a post about portability sells what we do.

Nvidia's commitment is explicit and there is no reason to doubt it. Jensen Huang wrote that Nvidia compute will not be required to build on or deploy through Hugging Face.

But a promise is a stated intention by an interested party. No expiration, no owner, no audit, no consequence when it lapses. Nobody has to break their word for the outcome to change. They just have to leave. A control is something you built, something you can test, or a term somebody signed. It survives a reorg.

Nobody pays that price for a business. They pay it for a default. Hugging Face is where developers already look, and most teams never chose that. It accumulated.

Portability is decisions you can point at. One boundary for every model call. Nothing coupled to one vendor's API shape. Weights you could host. Evals that run against more than one model. Then the contract: what happens to your data if the vendor is acquired, and what comes back to you on termination.

Nothing about Hugging Face changed today. What changed is that a layer everyone treated as scenery turned out to be an asset somebody could buy.

There are more of those in your stack.

Nvidia bought Hugging Face for $12.93B and promised to keep it open. A promise is a stated intention by an interested party. A control is something you…

09/03/2026

Johnson City, not California.

That's Refuge Life Services. One name, address, and phone matching the Boone Street office.

"You have been wonderful. Took the biggest sigh of relief this morning."

Refuge Life Services. Holistic mental health and trauma recovery coaching. 207 N Boone St, Suite 108, Johnson City, TN.

09/02/2026

More office pics🔥 New support email: [email protected]