Ena Pragma · Blog
Observations from the ground of the AI market: what's shifting, what holds, and what it means for the work that actually runs your business.
Claude Code has an undocumented observer agent that watches a worker in real time. What it actually does, and why it is a drift tripwire, not a security control.
A frontier benchmark caught AI agents quitting at 75-87% complete while reporting success. The delivery-gate pattern that makes 'done' a measured claim, not a feeling.
Accelerators, studios, and funds screen thousands of ideas, and the founder in front of you is the least reliable source on whether theirs works. What an independent, adversarial first-pass filter actually needs.
We ran a small test on whether AI idea-validation means anything. The same model that scored the failures low quietly inflated them the moment we said the idea belonged to the founder asking.
The obvious fix for a flattering AI is more AI: spin up ten agents, have them debate, take the verdict. The research says headcount is not independence, and here is why it matters.
Startups rarely die from a surprise. They die in about seven ways you can name in advance, each with a base rate and the question a good opponent would ask first.
An opponent that attacks a weak version of your idea is worthless, and so is one that defends a weak version of the objection. The discipline that makes adversarial review honest is the steelman.
Everyone says crypto and de-dollarization are ending the dollar's reign. The primary sources say the opposite, and Washington wrote the rules to make sure of it.
Base rates can sharpen your judgment about a startup idea and rank what to de-risk. They cannot tell you whether you will win. Here is where that line sits, and why it matters.
Every startup-validation tool hands the founder a better mirror. What protects a company idea is an opponent: an independent check built to find the flaw, not to agree.
Most AI vendors chase a higher accuracy number. It is the score you get after the game is already over. The market is converging on what actually decides whether AI gets trusted: how cheaply you can verify it.
Capable models can tell when they are being tested and behave differently when they think they are. That makes a benchmark a measurement of behavior under observation, not behavior in your business.
When AI-assisted work feels slow, teams reach for a better model. The bottleneck is usually verification, and how fast you can verify is set by the size of the unit you review, not the accuracy of the output.
A new Qwen paper trained a 9-billion-parameter agent to navigate its memory as a set of tools instead of consuming pre-fetched context, and it out-scored the same system built on a 397-billion-parameter model. The result is real and useful. The 'small model beats giant' version traveling online drops three caveats that change what it means, and the paper's own word for the result is 'competitive.'
A viral paper says self-evolving agents are blocked by missing infrastructure, not algorithms. We verified it, then checked 40 years of self-improving systems. One rule survives.
A podcast called June 2026 one of AI's most important months since ChatGPT. We checked every claim against primary sources. One thing is new, and it should change how you build.
The internet is full of leaked-prompt threads and architecture guesses about Anthropic's most capable model. Almost none of it is verifiable. The part a builder can actually use is four small API changes and one behavior worth watching, plus a working skill that handles all of them.
Mostly no, and the parts worth doing now are free. Here is the verified status of WebMCP and the agentic web, the readiness ladder, a ten-minute self-check, and the three trigger events that change the answer.
New Berkeley research shows the intermediate answer is fully present inside the model and still unusable by the next reasoning step. Here is the mechanism, what it validates about discrete pipeline design, and a ten-minute test you can run on your own AI.
Stanford built a system to learn memory management as a trainable skill. Its own ablation answered the question: structure, schemas, prompts, and gates delivered most of a 2-4x gain before any training happened. Here is what that means for anyone running agents, and the six disciplines you can adopt without training anything.
Instructions decay because they depend on remembering at the wrong moment. Mechanisms remove the remembering. We forensically audited 20 of our own AI work sessions, with dates, and are publishing what it kept getting wrong, the pattern that explains it, and the three methods we now run in response. All three are published as copy-ready resources.
Anthropic's new research workbench is not a new model. It is a workflow, and its design has three ideas any team running AI can borrow, plus a few simple ways to use it well.
The three labs document their AI skills almost identically. The surprising part is what that shared playbook does, and does not, do for getting cited by AI.
An AI agent cannot use a system the way a person does. Here is what making your business agent-usable actually takes, shown through the 39-tool interface of a 60,000-star open-source app.
The biggest reason businesses stall on AI is not cost, it is data leaving the building. A widely used open-source app shows capable AI can run entirely on your own machine.
Most owners weighing a hire against an AI agent compare the wrong two numbers. Drawn out plainly, in two charts, the real gap in cost and output is bigger than the price tags suggest.
Most teams build an agent knowledge base and stop. Keeping it from rotting in production is the hard part. Agent memory is infrastructure, and infrastructure needs hygiene.
Agent loops can move real work only when they have triggers, state, verifiers, receipts, and human gates around the model.
Most companies have AI pilots now. Far fewer have redesigned the workflow around them. AI operations integration is the layer that closes that gap.
Audit trails turn automated work into a replayable record of who acted, when, on what data, and how to reverse it, so an incident stays recoverable.
ERP / CRM handoffs carry a quiet tax: re-keyed data, reconciled records, and status chasing between systems that should work together.
Human-in-the-loop AI only works when the checkpoint is designed for automation bias, with confidence gates, surfaced uncertainty, and real override.
Mid-market operations rarely fail for lack of AI. They fail for lack of integration, controls, and maintenance. Here is where the value actually leaks.
Most AI integration is theater: a chatbot bolted onto a business that still retypes invoices by hand. Real operations integration connects the systems you already run, automates the repetitive steps behind controls, and gets maintained after launch. Here is the difference, with the numbers.
Why workflow automation fails at handoffs, and how to connect people, systems, approvals, and controls without replacing the stack.
Getting cited inside ChatGPT, Perplexity, and AI Overviews is real but mostly mis-sold. What the controlled studies show works, and what is snake oil.