A physical fleet changes the acceptance oracle, not just the testing tools

First-party operator account: Kuldeep Singh, “When Software Meets Hardware,” October 15, 2026. This retrospectively describes modernization of a retail fleet of about 800,000 battery-powered sensors across 3,000 stores. It originated in a submission to Thoughtworks’ 2025 Most Innovative Project; publication is not the deployment date. The team used AI to explore unfamiliar territory, prototype, and build dashboards, migration trackers and outage-detection tools; it neither isolates agent-written production firmware nor estimates an AI-caused quality or time advantage.

How the evidence environment changed. Historic docs did not reliably describe behavior. The team built a real sensor lab/farm, measured power and network behavior, and added on-device health/version/last-contact/update-state diagnostics exposed locally and remotely. A real store showed a device-count limit that small lab setups missed. Emulated hundreds of devices and virtual stores tested network/platform scale; real sensors still tested radio, firmware and power. Disconnection became expected because sleeping to save battery and reconnecting have costs. These are distinct oracle roles: simulations reach fleet scale; physical devices establish the conditions the emulator may omit. Across stores, aggregate alerts and dashboards let operations distinguish local from fleet failures.

Release and recovery. A remote firmware update consumes battery and can fail while a device is disconnected; rolling software back cannot restore consumed energy. The team designed secure downloads, validation, partition swaps, resumable upgrade state, and staged groups/stores with observed progress. Its account does not give rollout dates or counts, incident rates, active human hours, unchanged control stores, maintenance bill, or independently confirmed retail-operator value. It is a practical non-agent-specific counterexample to assuming that a green software check—or a green simulated environment—is a production pass. Compare FastyBird’s agent-built HomeKit repair: owner-observed device behavior and an explicit narrowing of unperformed trials qualify one issue, not all clients or long-run reliability. See release-oracle brief and prototype versus production boundary.

Hold: long first-person essay, possible weekend counterpoint if Dru engages physical acceptance or a current code-to-field trace; do not market this as a study of coding-agent effectiveness.