Weekly intelligence brief

Subscribe →

Weekly Digest

Edition №40 · 28–04 Oct 2026

Design, product, and travel intelligence. One brief, every Monday at 08:00 IST.

The week in five

  1. 1
    Apps, Agents, and Aggregation

    Computer use lets an agent work any app, API or not, so the screen stops being where a product wins its user. What still counts is being the service someone names when they tell the agent what they want.

    Product

  2. 2
    Kayak Co-Founders Launch Lola — Booking Holdings’ New AI Assistant: Exclusive

    Hafner moves the job from "what's the best flight" to "what should I do this weekend". That question comes before any destination exists, which is upstream of every fare grid Kayak ever built.

    Competitor Intel

  3. 3
    The New Big Ball of Mud: Why Agentic AI Systems Turn Fragile

    The proposed test is short. Could you say which kind of context to change, global, local or ambient? If not, the AI is the sole navigator of your setup. Keep global context brief enough to read in one sitting.

    Design

  4. 4
    Indian travel demand stays resilient despite global shocks: Booking.com’s Santosh Kumar

    The headline overstates it. The figures count searches on one platform while it spends more on Indian marketing, so part of the rise is its own push. Treat them as intent until bookings or revenue follow.

    Travel Sector

  5. 5
    Vrbo to Double Commissions as It Plays Catch-Up to Airbnb

    The steepest jump falls on PMS-connected managers, and AirDNA counts them at 55% of Vrbo listings worldwide. If the reported guest-fee cut lands, Vrbo is trading supplier margin for a lower price at checkout.

    Competitor Intel

Patterns & bets

Prediction ledger →

The pattern

For three weeks the design signals have turned on the gap between passing a check and being right. In W38 a typeface passed its automated spacing checks with a wrong lowercase 'a', and in W39 AI handled repeated work while a new module broke the design system. This week the fixes got practical: rules a script can test move into lint, and agents log what is missing so a person decides.

Open bets we’re tracking

Does Meta report how many Muse users have actually run tasks on their provisioned machines, as distinct from app downloads, before the end of the year?

Product — Apps, Agents, and Aggregation

Does Expedia Group answer with a separately branded assistant, or keep its AI planning inside the Expedia app?

Competitor Intel — Kayak Co-Founders Launch Lola — Booking Holdings’ New AI Assistant: Exclusive

Do MakeMyTrip's next quarterly results show domestic hotel growth near Booking.com's 20% search rise, or a gap that suggests the searches are not converting?

Travel Sector — Indian travel demand stays resilient despite global shocks: Booking.com’s Santosh Kumar

By the numbers

16 of 17

Indian hotel markets tracked by STR that recorded higher weekend occupancy in 2026. Weekends were historically the weak days for city hotels.

The Economic Times ↗

Signal of the Week

Stratechery28 Sept 2026

Thompson asked Meta's Muse to organise his saved Instagram recipes and had a working app in five minutes, running on the virtual machine Meta provisions for every US user. He argues agents reduce every app to an implementation detail, and that Muse and Microsoft's new Copilot lead because one reaches nearly every person and the other nearly every employee.

Pattern

On 25 September Thompson framed Muse's fight with Amazon as an aggregator war over shopping. Three days later Microsoft's Copilot shows the same design, an agent with its own cloud computer. The contest he now describes is over the one agent most people keep.

“What is happening with agents is that the ability to do stuff is becoming abundant; what is scarce is volition.”
— Stratechery

Design & Product

A rule written for an agent sets its path, and the agent still drifts. Evil Martians moved every rule a script can check, such as a ban on raw hex colours, out of the agent's instructions, and had agents log missing components in a gap journal instead of editing shared code. NN/g shows the cost of skipping that discipline: one participant lost 45 minutes after the AI pulled the wrong articles from a folder only it could search. The fix that holds is small. Keep each task precise enough that drift shows, turn a mistake that repeats into a rule, and end every request with a short review step that flags what wandered.

Nielsen Norman GroupThe New Big Ball of Mud: Why Agentic AI Systems Turn Fragile

In an NN/g study of non-engineers building agent workflows, Tanner Kohler found five of the patterns Foote and Yoder described in 1997, from untested rules files that become load-bearing to a junk-drawer folder only the AI can search. One participant lost 45 minutes fixing instructions after the AI pulled the wrong articles from such a folder.

Why it matters

The proposed test is short. Could you say which kind of context to change, global, local or ambient? If not, the AI is the sole navigator of your setup. Keep global context brief enough to read in one sitting.

FigmalionAI makes design system guardrails mandatory; this framework delivers them

A framework from Evil Martians splits agent work into two workflows, one for building screens and one for changing the design system, so an agent that finds a missing component logs it in a gap journal instead of editing shared code. Rules that can be checked mechanically, such as a ban on raw hex colours, move out of the agent's instructions into scripts and lint rules.

Pattern

Second write-up in a week on design systems written for coding agents, after the AGENTS.md approach on 1 Oct made code the source of truth. This one adds enforcement, because agents drift from rules they only read.

Creative BoomI helped shape McLaren's brand. Here's what worries me about the refresh

A former global head of brand, who worked on the new identity before leaving at the end of 2025, says the backlash goes deeper than a logo. He argues the fashion-house styling borrows the look of Chanel and Tom Ford without the conditions that made it work for them.

Pattern

Logged here on 24 September as a heritage refresh that kept every recognised symbol, following Instagram's in August. Keeping the symbols has not spared it. The objections land on the luxury styling around them, a part the wordmark-only reading missed.

Travel & OTA

The fare grid is the second stop. A trip in India mostly starts with inspiration on Instagram, and the fare check comes once a shortlist exists. Booking Holdings' new Lola assistant goes after that first step, starting from "what should I do this weekend" before any destination is picked. The weekend demand is real: STR counts higher weekend occupancy in 16 of the 17 Indian markets it tracks. Whoever answers that first question decides which hotels get named, and AI assistants today quote the Booking.com rate while the hotel's own price is usually absent.

SkiftKayak Co-Founders Launch Lola — Booking Holdings’ New AI Assistant: Exclusive

Kayak co-founders Steve Hafner and Paul English built Lola in five months as a wholly owned Booking Holdings startup, with about 20 staff, fresh code and its own brand. It aims to take a traveller from an open-ended idea to a booked trip in one chat, backed by a membership model and human concierge service.

Pattern

On 23 September Fogel said building an AI travel agent "wouldn't be that hard" and that trust was the harder part. Six days later the group launched one, built by Kayak's founders and backed by a human concierge.

LivemintIndian travel demand stays resilient despite global shocks: Booking.com’s Santosh Kumar

Searches on Booking.com from India rose through January to August 2026: domestic accommodation nearly 20%, domestic flights 25% and outbound flights 24%. The company is localising payments with UPI and RuPay while expanding marketing and rebuilding supply.

Pattern

Booking.com's two earlier India releases were surveys: 96% would adapt a trip rather than cancel (August), and 35% book within a week (September). This one draws on its own search data instead.

SkiftVrbo to Double Commissions as It Plays Catch-Up to Airbnb

From 29 October, Vrbo moves hosts and property managers to a flat 12% fee, up from 5% for PMS-connected managers and 8% for other hosts. A vacation rental expert told Skift that guest fees, now 11-14%, are set to fall sharply.

Pattern

Second rental-platform fee rework in a month, after Airbnb began testing lower fees for hosts who bring their own guests. The split between what hosts pay and what guests pay is back in play at both big rental platforms.

Signals worth keeping

“What is happening with agents is that the ability to do stuff is becoming abundant; what is scarce is volition.”

“At Lola, we're trying to solve for what should I do this weekend.”

“A note to hosts announcing the update did not address guest fees, which currently sit at 11-14%, but a vacation rental expert familiar with Vrbo's plans said the company plans on drastically reducing them.”

“AI is fragmenting travel distribution across many starting points instead of simply replacing Google with a new gatekeeper.”

“Under the proposed "list once, reach multiple platforms" model, tourism providers will be able to publish their offerings once and become discoverable across participating applications.”

“The guest who asks ChatGPT what a room costs at a specific hotel receives the Booking.com rate.”

Editor’s note

Most Indian trips I see start on Instagram. A hill station or a weekend getaway near the city goes on a shortlist first, and the fare check comes second. Even then it depends on who is going: younger travellers lean to the bus or the train, others fly. Around the last long weekend my feed filled with posts about long queues at the famous spots. My hunch is that AI is part of the reason. Anyone leaning heavily on an assistant to plan gets pointed at the same few places as everyone else. I cannot prove it yet. When an assistant plans your long weekend, does it know about the queue it is sending you into?

— Paarth

Get each new edition by email

Prefer LinkedIn? Follow Signect →