Netflix’s small-cohort AI Discover TV trial gives each capability declared inputs, outputs, context access and failure behavior. A controller checks approved steps, input producers, acyclicity and exactly one final page; invalid or slow plans fall back to tested plans. Sensitive member values enter as typed placeholders only after validation, while code enforces content settings. Taste summaries are precomputed, and plan records support replay. This is a themes-and-tuning trial, not free-text chat.
Harvey’s Joey Wang describes a different scaling failure: a release bug produced workspace-by-workspace alerts, launching roughly 900 agents on the same problem. Joining alerts to existing investigations reduced duplication, but hundreds of legitimate tasks still required a fair queue, per-person concurrency limits, interactive priority and capacity borrowing. Unattended automations are disabled after five consecutive failures. Local reusable repository copies reduced checkout costs; rebasing and CI against a moving main branch remain a merge-queue bottleneck.