Claude Managed Agents test: nightly review patrol needs ID tracking

The test used two public restaurants as samples: Katz's Delicatessen on the Lower East Side, with a 4.5 rating and 56,460 Google reviews, and Haidilao Hot Pot in Flushing, Queens, with a 4.8 Google rating and 8,792 reviews. The nightly cadence is driven by volume: Katz's latest 100 Google reviews covered 3.97 days, or about 24.7 reviews a day, while the latest 50 covered only 1.80 days, with 29 posted in the prior 24 hours. A weekly job would need four Google pages per store and would deliver its oldest low-star review six to seven days late, whereas a nightly job reading 50 reviews covers the prior night with a day of headroom.

The cost picture splits between the API and the model. A single store's four nightly calls cost $0.005556, but model token fees dominate: with Sonnet 5.5, estimated model cost is more than 10 times the API cost. The recommended setup therefore filters raw responses down to 1–2 star reviews with a sandbox script before the model reads them, and uses HTTP with an environment-variable credential rather than remote MCP, whose 100,000-character response cap was exceeded by a 105,266-character Google response for 100 reviews.

Operational details shape the design. Scheduled deployments launch with up to 9 minutes of jitter, each run gets its own budget ($0.50 in the example) that is not a monthly cap, and missed triggers are not backfilled. Sandbox files are wiped between sessions, so the list of seen review IDs must live in a memory store, and MCP toolsets default to always_ask, which would stall an unattended night session. The test also found that 6 of 10 low-star Google reviews for Katz's were non-English and that all 10 owner replies were sent in a 22-minute batch, 9 of them to 5-star reviews.