How an AI agent silently broke a two-Actor data pipeline three different ways

A developer building an automated seller-risk assessment tool connected two existing Apify Actors via an AI agent using the Model Context Protocol, only to find the pipeline produced wrong answers despite showing no errors. The first Actor scrapes marketplace listings for seller tax IDs, while the second checks those IDs against public risk registries; keeping them separate optimises cost and scheduling but creates a tricky data-handoff problem. The agent's first wiring attempt passed a dataset ID between runs, which worked locally using a personal token but silently failed in cloud deployments due to insufficient cross-run read permissions. A subtler issue emerged from input schema design: the agent defaulted to whichever input field appeared first and had the most confident description, so the developer restructured the schema to promote direct value-passing and demote the dataset-ID option with an explicit permissions warning. The case highlights how agentic pipeline failures can be invisible at the infrastructure level, only surfacing when results are manually verified against a known source.