Eight software-delivery use cases across Travel — Flights, Bus, Hotels, Trains, Sonar, Check-in, Figma and Vuln — now run as Ark agent flows. This is what they delivered, what it replaces, and what it saves.
From 1 September to 9 October, Ark agents ran the software-delivery loop across eight Travel use cases — 98% of runs engineer-initiated, most ending in a pull request a human only needs to review.
Every session, per use case, across the window: green = completed, red = did not complete (failed, stopped or still in flight).
Of 2,795 sessions, 2,358 completed (84.4%); the rest are a mix of failedstoppedstill running over five weeks — including sessions interrupted by the Sep–Oct control-plane / arkd reachability issue (being fixed). Flights 85%, Bus 89% and Sonar 99% lead; Figma ran clean; Hotels 76%, Trains 72% and Check-in 74% sit lower, and Vuln (low volume) at 40%.
From a pipeline per service and a human clicking Build, to one gated train — an ephemeral stack brought up, automated regression run against it, then torn down — that records what shipped and explains itself when it breaks.
When a stage fails, only the logs that bear on it are gathered; a model reads them, finds the one thing that broke, and posts a verdict with proof to Slack — every claim quoting the log word for word, or saying it could not tell rather than guessing.
Eight use cases, 2,795 sessions, 84% success across five weeks — and the marginal cost of the next use case is days of config, not a platform build. Agents delivered 1,883 unique PRs this window, 98% engineer-initiated; engineers spent their time reviewing, not scaffolding.