Insights

Insights

Evidence we find interesting. Commentary we think is useful. Updated as things happen.

What this is

We track what's actually happening with AI adoption — not the marketing, not the hype, but the evidence. When we find something worth knowing, it goes here. Short, sourced, and always with our take on why it matters for your business. This is the feed that proves we're paying attention.

· Updated 31 Aug 2026 — integration live Observation

Loop Engineering: Turning Recurring Work Into Self-Correcting Loops

We built a loop-orchestration layer on top of our agent mesh: recurring workflows that verify their own output against evidence, learn from corrections, and keep the human on the authority path rather than the execution path. The hard part was not the automation — it was making verification enforced rather than assumed, and making the learning layer gated rather than runaway. We validated the design against three independent adversarial reviews before it stood — and two days later it runs: a 1,600-line runner with 23 passing tests, the first loop registered on our live cron fleet in supervised pilot, and the kill-switch already proven in a live halt.

Why this matters

This is the pattern behind most "autonomous agent" claims, made safe. The insight generalises beyond our stack: any team running agents on recurring work can ask the same three questions — what is the memory, what actually triggers a run, and how do you verify the output against evidence, enforced rather than assumed? Get those right and the human's job becomes supervision. Get them wrong and you have a faster way to ship mistakes.

Read the full post 23/23 tests
Observation

ThreeUI Shows What Agent-Ready Design Tools Are Starting to Look Like

Meng To's new open-source Three.js library ships every component with explicit "copy as prompt" and "copy as source" workflows. The Community tier gives teams 50 parent components and 160+ variants under MIT — designed so an AI agent can take a natural-language description and iterate the visual language without starting from scratch. The interesting part is not the 3D — it is that ThreeUI makes visual components legible to agents. We are running a controlled evaluation against two real surfaces.

Why this matters

This is Proof Over Promise in miniature. The copy-as-prompt pattern lets the agent own the routine of boilerplate and first-pass visual work, lets a small team expand its capability surface without immediately hiring a Three.js specialist, and lets the human keep choosing the dependencies — including when not to use ThreeUI. That is exactly the "Own the routine / Expand your team / Choose your dependencies" thesis in action, applied to a single library release.

Read the full post 50 / 160+
Research

Only 4 out of 500 S&P 500 Companies Score 100 on AI Adoption

The AIDE Index — an MIT-linked, objective scoring framework — rated every S&P 500 company on actual AI adoption. The result: 496 companies didn't make 100. Stated AI strategy turned out to be nearly uncorrelated with real implementation. Executives were 3.6× more likely to say "innovation" than "AI" on earnings calls — a telling signal gap. And the top 10 wasn't dominated by tech giants: Walmart, SLB, and AES proved that non-tech companies can lead when they commit.

Why this matters

This is the strongest evidence yet for "Proof Over Promise." If 99.2% of the biggest companies in the world haven't fully adopted AI — and the ones talking about it most aren't the ones doing it most — then the gap between stated ambition and real capability is the defining problem in AI right now. We built our business to close that gap.

Read the original research 4/500
Observation

We Pointed an AI Design Tool at Our Own Site — Here's What Came Back

We gave Claude a brief, then a design tool (Fable 5), then our own site, and asked it to show us directions we hadn't considered. It returned ten distinct concept treatments in a single session — different structures, different moods, different interactive patterns. The useful part wasn't the designs. It was the moment we asked it to tell us what questions we should have asked it in the first place. That reverse-prompt trick changed how we think about working with AI on creative work.

Why this matters

AI is most valuable when it shows you a question you didn't know to ask. The ten concepts aren't the point — they're evidence that an AI design tool, given the right brief, can surface directions a human team would take weeks to reach. We've published them as a separate lab so you can see the range for yourself. This is exploration, not a finished product — but it's the kind of work we do with our own brand before we'd ever do it with yours.

Open the Concept Lab 10 concepts

More insights on the way. When we find evidence worth sharing, it goes here. No schedule, no filler — just signals.

Proof Over Promise

Want AI that actually works?

These insights shape our approach. If you'd rather see proof in practice, our case studies show what happens when evidence meets execution.

See Our Case Studies Start a Conversation