Developer builds LLM-powered media manager, learns hard lessons about constraining AI agents

A developer built a self-hosted tool called Mediary Scout that uses an LLM agent to search, transfer, and verify media files on cloud storage drives. During testing, three critical bugs emerged from giving the AI agent write access to the cloud drive, exposing the risks of letting language models control irreversible file operations. In one case, the agent redundantly queued eleven overlapping season packs across sixteen searches, exhausting the cloud provider's API rate limit and killing the entire acquisition run. The developer found that prompting the model to show restraint was ineffective, and the real fix required deterministic code — including a set-cover algorithm to minimize redundant transfers and a hard cap on search queries. The core design lesson was that LLM agents should only propose actions, while surrounding deterministic logic enforces limits and executes all irreversible side effects.