Why wrapping every API endpoint as an agent tool can backfire

The article argues that exposing each backend endpoint as a separate agent tool creates shallow interfaces that push multi-step logic onto the model. Such designs inflate context with tool schemas, compound error rates across long chains, and leak business rules into prompts that models may ignore. The author recommends deeper, task-shaped tools for customer-facing automation that handles money, with invariants enforced in code. Cited evidence, including a non-peer-reviewed Meta preprint, is described as suggestive rather than a controlled comparison. The piece suggests merging frequently sequential tools and moving prompt rules into tool-level checks as practical steps.