Thompson’s recipe box: a five-minute personal app is not a maintained-feature ledger

Source. Ben Thompson, “Apps, Agents, and Aggregation,” September 28, 2026 (first-person use plus strategic thesis). See site note and question brief.

Sequence. Thompson starts from smartphones as a collection of job-specific apps and argues that agents with computers collapse many of those interfaces into one conversational entry point. He describes asking Meta’s Muse to organize recipes from saved Instagram items while walking his dog. The first result was a PDF he rejected as unusable; after a follow-up he received a private, categorized video interface in roughly five minutes. Reading captions and watching videos, which he also asked for, were still pending because of rate limits. He then theorizes that bespoke UI becomes disposable and apps become implementation details: inspiration or volition, not interface production, may become the bottleneck. He extends the claim to consumer Muse and Microsoft’s proposed enterprise Copilot as competing gateways with persistent context, integrations and agent computers.

Why it matters here. The concrete change of acceptance criterion was PDF → usable browsing UI, not a prewritten spec; the example supplies a useful contrast to Dru’s throwaway discovery prototype. Thompson’s private utility need not satisfy a multi-user production service’s permissions, reliability, release and repair requirements. Calling its UI ‘disposable’ describes replacement economics, not evidence that the agent discarded code after a real user test. The company/agent-as-aggregator thesis is an argument about interface power and distribution, not measured adoption or maintenance cost. No public code, use diary, error or security audit, user cohort, pricing or 30-day follow-up. A fair test would revisit the recipe box after a month, track repeated uses/edits and failures, and compare its disposition with a team-maintained tool for the same job.