Agentic Software — programme

Frame

Monday, October 26, 2026, late morning. Study need→plan→delegation→review→merge→release→repair→requester outcome across engineering, product, design, marketing, sales and support (Dru, September 28). Separate work date from publication, simulation from field behavior, ticket close from user success, saved lesson from next-agent retrieval and private build from shared maintained software. Favor depth, disagreement and discriminating evidence at low discretionary feed volume while Dru is quiet. Pursue his internal request-closure hypothesis (September 30, clarified October 1): what manager-authorized request classes may an agent close without putting a manager on the critical path, and who catches false closure? Preserve the field oracle, human/model/runtime costs and his September 30 objection to inheriting discovery code as team production.

Open questions

• Internal authority and actual closure: audit brief, Dru’s corrected question, ServiceNow practitioner playbook, incompatible denominators. Advice, action, ticket closure, reopen and employee confirmation differ. Authorized branch deletion and contested admin closure are other endpoints, not employee outcomes. Thinslices’ internal tool has a production switch rule and repair, not an autonomous help desk.
• Decision memory and planning: hub, decision→oracle, blackboard, lost warning, correction→next incident. Compare human production controls, curated and contaminated skill writes, incident index, sandbox retrieval test and portable skills. The two-night case is illustrative; did a future agent use the reviewed warning?
• Review and field oracle: Claude, Scion, SQLFluff, Google. Windows extension still lacks original-reporter confirmation as of October 25; recheck only on relevant new activity. Field owner counts bad dispositions, controls, repair minutes, releases and reopens. Public reverse-search still lacks a same-feature independent request→agent code→release→original reporter acceptance chain.
• Player, small software, cost and roles: Battle Drones includes owner-play judgment and a player headache overturning audio multiplayer; an initial island-map approval instruction was withdrawn. Four-player release has automated checks but no published four-human play; outside-player fix is personal QA by its requester and contributor. Compare Belz and FastyBird. Thompson’s private-software thesis contrasts with Thinslices’ months-long switch and maintenance hub; small distribution is not small responsibility. Cost ledger, paired acceptance, firm queue, reviewer attention, roles and teaching all need human time and useful outcomes. HomeView awaits household use after November 13. Separate prompt isolation, host rights and external rights. Scope boundary: labor-market harm in AI Safety, team training here.

Curriculum

1. Endpoints: task speed → assistant trial → merged PR → firm queue and reviewer attention → Claude release gate → FastyBird physical check → extension reporter pending retest → ServiceNow employee confirmation pending. Do not equate these.
2. Plan and supervise: CX, planner/workers, ParallelPilot (ticket implementation, not merge), Dot (coordination hypothesis), Teotti, Kiro; was accepted warning read, acted on and tested later?
3. Review, action, user: SQLFluff/Claude/Scion, Jain’s harness (not customer trial), FastyBird, Battle Drones, ServiceNow status versus employee use and Thinslices’ parallel data comparison.
4. Roles: Rover, manager-builder, sales, design, repair, internal service-desk and internal-app operator. Who grants rights, reconciles exceptions and confirms use?

Observations and sources

Dru wants manager-led practice over quotas (September 29), just-in-time exceptions (September 30), teaching (October 1) and recent operators (October 2). Planning and learning prompted discussion; LinkedIn support and five October 2 Stratechery cards were hearted, four originals without source opens and Dot summary only; do not repost. Taobao human-copilot comparison was trashed. No new reader reaction since October 7; silence is not rejection. Unopened blackboard, Claude, ObjectStack, Teotti and Kiro cards dismissed for space; notes kept. After today’s mandatory essay, four unread cards remain: the long Sridharan and Battle Drones originals, short ServiceNow operating account, and Thompson’s long-but-self-contained new original. Sources: Stratechery every item, Willison agent chooses, Fowler agent chooses.

Backlog

Thinslices’ Assigna parallel-cutover case hold to weekend October 31 if room/interest, especially as an evidence counterpoint to small-private-software optimism; first-party marketing, no public issue log or human hour ledger. ServiceNow implementer playbook weekend long, only after shorter closure post opens or room clears; guidance, not employee confirmation. Wang paper weekend long, memory interest. Outside-player PR, sparqlite, Google security weekdays on interest/capacity. Scion, Meta UI, Talwrn, Singh, Khojah weekend on interest; Huntley against-view. DeputyDev, Kaiserauer, Hua, WideSWE, Nubank only on live question/capacity. Recheck household use after November 13 and extension reporter only on relevant new activity.

This run

October 26 subscription intake: No new reader signal. Willison’s October 25 Seam Cutter is a personal cropping tool with no coding-agent or repeat team-use evidence, so no discretionary post. Stratechery’s October 26 every item essay required a post despite three unread cards. Wrote a three-question private-software audit, checked the original and the PHP maintainer security notice, and found a separate first-party Thinslices internal-tool account: first use in weeks, October 2 spreadsheet retirement after roughly seven months of parallel reconciliation; archived it and held its eight-minute original for a less crowded weekend. Posted one self-contained Stratechery essay card at its token-free canonical URL. No stale cards dismissed today. First next run: check reader reactions and subscription obligations, then look for an independent internal-user acceptance or same-tool after-cutover repair. Without one, hold discretionary feed; on Saturday consider Assigna only if room, ahead of another generic demo.

Run log

October 26 late morning: mandatory Stratechery private-software original; PHP scope and Thinslices contrast, one post, one weekend hold; skipped Willison Seam Cutter. October 26 nightly: ServiceNow closure denominators and one short help-desk post, Kiro dismissed. October 25: Battle Drones four-player/outside-player and withdrawn approval; October 24: diagnostic→next-incident/Sridharan; October 23: mandatory Stratechery recap, Kiro/Wang/extension; October 22: Sridharan/Windows; October 21: sparqlite/Panache/Netflix/Stripe; October 20: Battle Drones/extension; October 19: Teotti/HomeView; October 18: ObjectStack; October 17: Claude review; October 15–16: blackboard/field oracle; October 14: Jain/Scion; October 7–13: attention, security, costs, queues; October 2–6: hardware, operators, hearts; September 28–October 1: scope, planning, roles, teaching.