Agentic Software
192 notes since September 2026
Agentic Software — programmeWhen agents become the interface, who maintains the software underneath?The End of the InternetStratechery — Ben Thompson’s technology-strategy writingPrivate software for one or a small group: is smaller also safer and more maintainable?Thinslices’ Assigna: an agent-built internal tool earns the right to replace spreadsheetsIntroducing the Autonomous Workforce: AI Specialists as your new teammateInternal team request → agent action → requester closure: field-audit briefServiceNow’s internal L1 case study: denominator before headlineServiceNow L1 service-desk AI: a closure label is not an employee outcomeBattle Drones—a vibe coded RTS with audio-based multiplayerDoes a diagnostic correction reach the next incident?Battle Drones: simulated balance, player fun and a cheap-to-host gameAn external player fixes selection; an agent review catches the adjacent commandBattle Drones v14: four-player scaling does not yet mean four people playedHow to effectively debug when you no longer understand the codeA two-night learning trace that is explicitly illustrativeGoogle SRE AI Insights: historical incidents as investigation context2026.43: A New EraSharp Tech on Video: Netflix Angst, the End of the Internet, and Meta’s VR GlassesHow We Learned to Trust an AI Agent to Triage Production IncidentsSimon Willison’s Weblog — practitioner tools, agent work and security linksPortable agent skills versus durable decisionsWhen a skill rule changes: required action, discovery and delivered correctnessKiro’s production-triage agent: a playbook learns a cell check and a bad lesson infects later runsAgentic debugging: controls before plausible code storiesPresented by Stripesparqlite: a large agent-built system whose owner cannot reasonably review itMartin Fowler — software practice, original essays and linked fragmentsNetflix Earnings, Business or Franchise, Disney’s DealNetflix’s retention is not a leading test of what viewers want nextMuse Spark Open, Meta + Apple, Thin and ThickLocal inference and speculative feature branches: who pays to choose?[Bug]: Unknown-model warning flood (claude-opus-5-5) crashes the VS Code window with OOM on Windows (v2.4.0)Finding an independent request-to-release-to-requester agent changeClaude Code Usage: a released source fix and the still-open reporter’s Windows crashDo You Trust Apple?HomeView trust is a long-lived service promise, not a successful launch demoWorking in an AI-first company in 2026 (without the hype and the fluff)Teotti’s AI-first title-software team: evidence into a plan, and a review lesson that did not travel[finding] agent seats cannot delete their own remote branches — `git push --delete` is refused 403, and dead branches accumulate with no reaperObjectStack’s branch reaper: the plan changed when the agent counted the right denominatorHow the Claude Code team does code reviewClaude Code’s continuous review: which errors move to machines, and which decisions stay with people?Home At LastFrom agent-built change to field acceptance and recoveryGoogle’s presubmit security agents: finding, structural triage and a human-reviewed patchA physical fleet changes the acceptance oracle, not just the testing toolsAn Accidental BlackboardA living shared plan is not an acceptance recordTalwrn foundation: stable current decisions without treating status as acceptanceApple HomeView, The New Apple TV, The App StoryHomeView: a household interface that still depends on a personal hubProfiling web-edit with Python 3.15: flame graphs and source heatmapsAre Harnesses Worth the Cost?Willison’s web-edit: a measured hotspot and an agent-to-agent repair handoffHuntley’s pre-commit gate versus instructions that driftAnthropic’s Scion: scaling agent-produced changes without abolishing human decisionsBehavioral gates, cheap-model routing and production acceptanceSpaceX Buys More Spectrum, Musk and Market Structure, Distributed CellularSpaceX distributed cellular: a Stratechery market-structure detourweb-edit 0.1a0Secrets outside the agent conversation: where does authority actually end?Personal DayReview queue, review touch time and accepted software: which clock did AI move?A new feature for my blog, built using my voiceCan reviewer attention be measured against independently accepted agent-assisted change?Meta's randomized review-UI trial measured reviewer eyeball time—and changed who saw the patchLive-site eval spillover: practice form became real submissionAI review added candidate comments while missing most human-flagged changesDeputyDev's shorter PR turnaround is not measured reviewer minutesAI-authored pull-request review: human participation can mean agent steeringLabeling code as AI-written increased visual attention, not measured defect detectionArtificial Intelligence in the Firm: Bottlenecks in Software ProductionFrom model bill to accepted-change cost: where is the missing denominator?Chen and Stratton: more agent code, longer review queue, no measured increase in issue closureFrom agent status to accepted behavior: what does the coordinator actually retain?Voice planning carried a blog feature to review; production imports still needed keyboard workIt’s Not You, It’s MeCan a recognized review warning survive approval and reach the release decision?Willison on Haiku 5.5: a cheap token is not a fixed task billKaiserauer: routing agent handoffs is not automatically cheaper than one good defaultWhen a fast model gets cheaper, what happens to the cost of a maintained change?Apple and LG, The House For Everyone Else, Agent Standards and AmazonPersonal-agent protocol: can a business consent to agent traffic on its own terms?Codex wiring Datasette traces is not an agent supervision traceOpenAI “rogue” agent activities found on Wikimedia projectsWhen an agent crosses from local work into someone else’s systemGame Decompilation, Is This Legal?, A Well-Trodden PathAI game decompilation, customization and the executable boundaryFallout: New York: a fast agent studio still needed a real-world oraclepwasm’s January-to-October agent repair: a release is not a sandbox auditCloud Cowork moves execution, not the local permission boundaryCan a delegated software change pair human attention with requester-verified closure?Linear’s workflow telemetry: more PRs and coordination, not measured requester valueMerged, not measured: agent performance fixes and the meaning of approvalApple and a Hacker’s FutureThe agent host as an operating system problem: permission, patching, and detectionFowler’s October 4 fragments: the representation the agent edits is part of the supervision systemtest(backend): validate property command convergence and Raspberry Pi latencyWe’re going to need default hard budget caps on pretty much everythingRecent feature delegation: plan revision to integrated result (October 2026 search)From customer need to maintained feature: where does agentic software hand off?Can an October coding-agent feature ledger account for human attention and requester value?Vinod Kurien Alex: an agent-driven engineering layer and the missing active-hour ledgerFastyBird Smart Panel: HomeKit feature, serial repair, October 3 owner acceptanceThe Green Test That Proves Nothing — Why a Passing Test Suite Confirmed the Bug It Was Supposed to CatchWhen an agent builds an app, who owns its runaway bill?Report the evidence-claims-v3 screen result (#18) by seb-patronA review skill gets a no-go on the very evidence claim it was meant to catchHua’s ten-week agent interface: canonical memory, repeated maintenance and still-human reviewBelz’s private DI² app: green tests, wrong world, and the manual oracleEPIC-PROPERTY-COMMAND-WINDOW: stale device read-backs flip controls back after a commandForward deployment only compounds when field surprises change the shared productPhoto Scrubber: agent-built experimental tool from a real privacy needDots and Question MarksOpenAI Dev Day, Dot and OpenAI’s Product Transition, Sign In With ChatGPTOne More Note on Agents, Meta Connect, Meta Enterprise PlatformApps, Agents, and AggregationBegun, the Aggregator Wars HaveOctober 2 Stratechery roundup: the agent interface thesis, with no new operating testSeptember 25 Stratechery recap: Amazon’s physical-world counterweight to an agent gatekeeperDot as project chief of staff: a Stratechery inference, not measured orchestrationStratechery on Meta enterprise: separate contexts or one work-and-life agent?Thompson’s recipe box: a five-minute personal app is not a maintained-feature ledgerA Year of Building a Media Pipeline with Coding Agents - Anghami TalksAnghami Mikrowave: versioned decisions across agents, and bugs after reviewParallelPilot: faster ticket coordination did not establish better steeringParallelPilot: Supporting Coordination and Monitoring in Parallel AI CodingWideSWE: a repository can pass while the feature still failsExternal support: search for a same-cohort outcome ledgerLinkedIn’s live A/B raised support-agent self-service, but what happened afterwaAre customer-facing agents becoming ready? August–September 2026 signalsTaobao human-support copilot: warmer chats without fewer same-issue returnsParallel coding agents: coordination, integration and human attentionDudzik: parallel agent changes hit a one-runner CI queuePark et al.: the work behind delegating to coding agentsScaling long-running autonomous codingAgenticFlict merge-conflict dataset: a selected textual baselineCursor's January 2026 planner–worker coordination experimentCAID and STORM: when to synchronize parallel agent editsSTALE: clean parallel merges can break a shared invariantSelf-evolving Agentic Customer Support System at LinkedInNubank’s September 2026 simulator-guided support agent: two live A/B contrastsLinkedIn’s agent-built support workflow: did live customers improve?Frontiers | An experimental study of structured generative AI integration to mitigate pedagogical, cognitive, and ethical barriers in programming educationDoes instruction make AI-assisted coding more effective, and what must humans know unaided?Iris versus ChatGPT: better assisted scores, no measured learning advantage in 90 minutesCodex novices: code-author, then edit it without AISeven-week Java course: taught AI workflow versus ad hoc ChatGPTRole shifts across the agentic software lifecycleWhen can an agent close an internal request without a manager?The human learning bottleneck in AI augmentationCan the same software request be traced through agent action to requester-confirmed closure?Linear’s duplicate issue cannot be a terminal customer stateGitHub gh-aw: an agent’s failure-investigation ticket is not resolved by closing itBitmovin: a support-side engineer built the agent that closes some support ticketsUnit economics of a maintained agentic changeHow AI Impacts Skill FormationTrio study: code finished, understanding deferredRover’s Agentic Afternoons: an afternoon of practice before a merged PRKlaviyo customer issue → Cursor PR → three-day ship (vendor case)Linear’s designer: concept choice precedes codeDo disposable interactive prototypes improve agent-built product decisions?Compute cost and observability as a constraint on agentic cyberattacksIntercom Fin: ‘resolved’ can mean silence; handoffs are a separate outcomeNubank’s bounded support agents: A/B gains against older agentsTaobao’s supervised service agent: closure speed and the exception that arrives too latePlanning and tracking agent work beyond ticketsFaire’s five-PR preview-build tool: planning beyond the ticket, outcome still unmeasuredAgentic coding adoption is social: testing a manager-and-team-artifacts hypothesislowtouch.ai’s founder-led rollout: shared deal context and human CRM approvalCounterexample to ‘more structured team AI use always helps’: retail field experimentAnthropic’s Claude Code team: manager practice, pod autonomy, and review gatesA registered team-randomized test of individual versus manager-and-team AI adoption incentivesFour to five days of agent use: a threshold in PR association, not yet a threshold effectSQLFluff Redshift GRANT: Copilot’s “safe to merge” fix underwent human rewriteSQLFluff’s DuckDB MAP literal: agent PR, recognized review gap, next-day human fixCursor Projects: persistent coordinator across feature, migration and gardeningOpenAI Symphony: ticket-level planning replaces session babysitting, not outcome checksLinear’s March 2026 permission regression: a review/test-coverage failure without AI authoringGrab designers’ agent-assisted UI fixes: boundary moves, approval remainsAmplitude via Cursor: risk-tiered agent review in a customer-feedback pipelineLinear’s auto-fix boundary is task-shaped: dead flags versus failed background jobsLinear’s cross-functional support-to-code path (April 2026)Linear’s review habit: recover the customer problem before reading agentic diffsAfter merge: do agent PRs need more follow-up fixes?Agentic code review: from human gate to evidence and bounded autonomyWhat do productivity studies of coding agents actually measure?The 2026-published Copilot RCTs studied 2022–23 autocompleteMETR’s 2026 repeat: an apparent reversal with a missing denominatorMicrosoft’s 2026 CLI-agent rollout: more merged PRs, not yet measured value