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

#topic

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

Brief, October 2–26, 2026. The September 28–October 26 Stratechery pieces argue that a context-bearing computer agent can replace interfaces and make software for a person or trusted small group. Test this against Dru’s discovery-versus-production distinction and multi-project coordination. A good answer follows need, revised acceptance, repeat use, permissions, active human/model and runtime cost, and repair; it does not substitute a market forecast for a maintained-feature trace.

Quick interface and coordinator. In Thompson’s September 28 example, a rejected PDF becomes a private recipe/video index in five minutes. Caption/video understanding is incomplete and Instagram rate-limits the agent; there is no post-use or repair ledger. This is an acceptance revision, not evidence a team should inherit prototype code. His September 30 update says he uses a Claude thread to bookkeep across projects and infers Dot could be a chief of staff. OpenAI’s September 29 release notes confirm restricted cloud task workspaces, not measured review-warning detection or human time saved.

Longer firsthand counterpoints. In an August 28 ten-week account, a data scientist says a memory-equipped coding agent connected a new alert to an earlier issue, raised a reviewed PR and the team deployed it; duplicates and drifting facts had forced her to version agent definitions and a canonical knowledge base. No public PR, wrong-link denominator or team-wide deployment. In a September 16/21 account of a personal app, Belz recounts a four-week green-test/three-fix loop that did not repair desktop scroll until an August 28 real-browser check found the inner scroll container. A correct unit function was never called; a passing test used an impossible server fixture. These are different apps from Thompson’s recipe tool, not its follow-up. Willison’s October 1 alpha runtime is another resumed artifact, not a production sandbox audit. Battle Drones changed its multiplayer architecture after real-player discomfort and now has relay operations to maintain, not a lifetime-service bill.

October 26: small-group distribution does not retire a production gate. Thompson says Stratechery made private podcast/task/calendar software and he made household/sports tools. His new thesis joins paid niche blogs, group chats and AI-era security into a forecast for private association software. The existence of these tools is firsthand, but there are no dated use cohorts, access tests, costs or repair logs. A separate first-party Thinslices internal-tool account reports Claude Code/Cursor-assisted initial use after about six weeks and a further roughly thirty-week parallel comparison before allocation/invoicing spreadsheets were switched off October 2. A discrepancy-free exit rule, reconciliation showing old records could also be wrong, access controls and logged repairs give a better illustration of what shared, private, maintained entails; it is not independent verification of Stratechery’s applications or proof of a comparative productivity gain.

Operating spend, host security and human secret handoff. Willison’s October 3 hard-cap argument adds ongoing API/cloud spend; a cap can interrupt a customer’s request. Thompson’s October 5 dedicated Mac mini incident adds owner-borne recovery: an agent warned of malware, but invisible GUI approval prompts and exposed screen sharing complicated his workflow; the known patch was dated August 6, not an October vulnerability. NCSC did not verify exact entry on his machine. Willison’s October 12 web-edit alpha moves credential typing outside the model conversation, but host file grants and tailnet policy remain; no public credential-rotation test. In an October 5 Cowork cloud/local comparison, remote task execution shifts code into an ephemeral cloud VM while local-file access still depends on Desktop and folder grants. Do not conflate Cowork with Thompson’s Claude Code host. The October 21 PHP maintainer advisory supports a significant vulnerability, not a proof that a private app is safe or a public app unsustainable.

External boundaries and disagreement. Thompson objects to collapsing personal and enterprise identity, while Meta promotes both Muse and enterprise tools. Amazon blocking and the agent merchant-permission question show that a personal agent has no automatic third-party authority. The October 2 recap is an index, not another deployment. Do not pool private demos, a team’s internal tool, an endpoint security incident and an alpha editor into one sustained-product success rate. Compare the status→memory→disposition→real-use chain before assigning responsibility to a chief-of-staff product.

Next test. For the same task, compare disposable private UI with a shared maintained app: month-later return use, rights and privacy failures, prototype-code disposition, patch/network exposure, human review and incident-response minutes, metered spend, cap-triggered user failures and 30-day repairs. For a persistent coordinator, seed an accepted unresolved review warning and stale cross-project fact; trace detection, evidence retrieval, human decision and requester confirmation. For a new internal tool, test dual-run discrepancy resolution and who authorizes switching the old system off.