"Type 'habit tracker', walk away." Pipelines like Hermes chain agents into a factory — but the interesting engineering is in the gates, not the glamour. Run the line below.
Modern pipelines reliably automate the mechanical middle: scaffolding a repo, writing CRUD code, wiring standard auth, generating tests, deploying. Classification and planning agents draft a spec from one sentence by pattern-matching against thousands of similar apps.
What they don't automate: deciding whether the app should exist, resolving ambiguous requirements ("habit tracker" — for teams? with reminders? privacy model?), and owning the consequences.
Toggle auto-approve above and watch the "unreviewed" counter climb. Human-in-the-loop gates exist because agent plans fail in confident, plausible ways. A 30-second human review at the plan stage is far cheaper than debugging a fully built wrong app. Rule of thumb: gate before irreversible or expensive steps (deploys, payments, mass emails).
Press "Inject test failure": the box returns to the Build station, not the trash. Good pipelines treat failed tests as feedback — the test report becomes new context for the build agent's retry. Bounded retries (2-3) with escalation to a human afterward prevent infinite loops burning tokens.
"Walk away and it ships" works best for well-trodden app shapes (trackers, dashboards, CRUD tools). Novel architecture, tricky integrations, and polish still consume human hours. Illustrative pipeline: ~5 stations here; real systems add code review, security scan, and staging bake time. Figures in this sim are illustrative estimates, not benchmarks.