Intelligence Track

Design

8 July 2026

The Brief

The most-quoted line out of the West's super-app race — that Uber owns the transaction and Google Maps owns the moment before it — buries the part that matters for travel: if the discovery engine's model wins, hotels become white-label inventory competing on price; if the transaction engine's wins, the same thing happens from the other side. Either way the room is the commodity and the layer above it captures the value, and that exact move repeats down the rest of the day. The country's largest hotel operator has stopped chasing ultra-premium international guests and is building mid-market rooms in tier-2 and pilgrimage cities where domestic demand actually sits, keeping the guest inside its own brand and loyalty rather than any listing; a product leader shows that the edge in agentic AI is not the model but the harness — the adapters, permissions and context scaffolding wrapped around it; and a designer argues AI has swallowed execution and is now reaching for the decision of what to build, leaving experience-forged judgment as the scarce input. Across all four, the raw thing — the ride, the room, the model, the execution — is being commoditised while value slides to whoever owns the layer that directs it. The open question for any platform sitting between supply and traveller: when the commodity below you and the model above you are both cheap, what layer do you own that neither can absorb?

Skift · 8 Jul 2026

Within a single quarter, a transaction engine (Uber, $52B+ 2025 revenue, 50M+ Uber One members) added Expedia-powered hotel bookings while a discovery engine (Google Maps, 2B+ users) layered a Gemini-powered 'Ask Maps' that recommends and reserves — each crossing into the other's territory, yet the piece argues neither becomes a true super app because the West's existing credit-card rails removed the embedded-payments wedge that built Asia's super apps.

Industry lens

Does 'Ask Maps' start completing bookings itself rather than handing off to partner inventory — the moment discovery stops referring and starts capturing — and if it does, do hotel chains pull inventory direct to avoid being commoditised inside it?

Reading as

AI & Design

What a harness is and how to build one with Claude Agent SDK

A working demonstration of the phrase 'it's not the model, it's the harness' — a Sentry bug-triage system built with the Claude Agent SDK, a custom Ink terminal UI, and opinionated adapters for Sentry, Linear, GitHub and Vercel that handles evidence-gathering, root-cause analysis and follow-up artifacts without a fresh human prompt each time — arguing the durable engineering work is the task-specific scaffolding, not the underlying model.

It relocates the defensible work in an AI product away from model choice — a swappable commodity — and onto the task-specific harness: the adapters, permissions and context logic that encode a company's own workflows, which means the payoff comes from building reusable harnesses around repetitive structured processes, not from prompt-tuning or picking the frontier model.

Lenny's Newsletter·8 Jul 2026

Design Ops

Also in Design Ops

Execution — the 'how' of building software — has been largely absorbed by AI tooling; the argument here is that decision-making itself, the 'what' to build, is the next contested territory, with judgment accumulated through lived experience and prior failure framed as the input that resists automation.