When agents become the interface, who maintains the software underneath?

Brief, October 2, 2026. The September 28–October 2 Stratechery batch argues that an agent with a computer and remembered context can displace app interfaces and build private, disposable UI. Test that against Dru’s discovery-versus-production distinction and multi-project coordination. A good answer follows user need and changed acceptance through deployment, repeated use, permissions, cost and repair; it does not substitute a business-model thesis for a maintained-feature trace.

Subquestions and results. (1) Observed personal UI versus generalized ‘infinite app’? In his September 28 firsthand example, Thompson rejects a PDF and gets a browsable personal recipe/video index in five minutes, but caption/video understanding is still being worked on, and no post-use or repair log follows. This is an acceptance revision, not a production quality result or evidence that discovery code was thrown away. (2) Does agent+workspace solve supervision? His September 30 update says he personally uses a dedicated Claude bookkeeping thread because Codex projects leave cross-project tracking to him; he infers Dot could fill that chief-of-staff role. OpenAI’s September 29 notes confirm separate cloud task workspaces and gradual Dot rollout, not measured cross-feature integration, review or successful redirection. Compare ParallelPilot and Anghami. (3) One identity or bounded access? September 29 argues work and personal contexts should be separate and that enterprise sales may divert a consumer agent; Meta’s September 28 announcement confirms the new platform and asserts privacy/security, but neither party reports an enterprise use/comparison ledger. (4) Do recaps add evidence? September 25 flags Amazon blocking a shopping agent and the Walmart alternative; October 2 is a guide to the week’s Meta/OpenAI arguments. Neither is an independent deployment outcome.

Disagreement and boundaries. Thompson sees one broadly connected agent as the likely gateway but argues against the same product/identity spanning consumer and enterprise. Meta promotes both Muse and an enterprise stack; that announcement does not resolve the strategic disagreement. His ‘apps are dead’ formulation is a prediction backed by one personal UI and agent-provisioning claims, not audited app disappearance. Stratechery’s September 25 roundup notes the supplier can simply refuse an agent access—a limit to frictionless abstraction. Initial source list: Monday article, Tuesday update, Wednesday update, October 2 roundup, September 25 roundup, Meta Muse and OpenAI release notes.

Next concrete test. For the same task, compare throwaway/private agent-generated UI with a maintained shared internal app: monthly return use, permission failures, user revisions, code/plan retention, human review minutes, model spend, and 30-day repair. For chief-of-staff products, trace whether a cross-project agent notices an accepted-but-unresolved review item before merge.