Agents as internal services: running an LLM agent inside the system, against the system's own tools

Most deployed LLM agents are conversational assistants attached to a product. This report describes the next step taken in one SaaS backend: using the same agent runtime as an internal service, invoked by a durable backend job with no user in the loop, to produce a text artifact that a second model consumes as context. The agent reads the system's own data through a tool gateway generated from the backend's OpenAPI description and exposed over the Model Context Protocol; a pinned mode fixes the tool grants, the preloaded skill and the output contract; and the backend treats the result as a build artifact, validated, stored where people can read and edit it, cached by a normalized identity, and evaluated with the consuming model held constant. Three runs against a human-written baseline showed that the first quality failures were wiring, caching and sampling problems rather than prompt problems, and produced two rules for context written by one model for another: absence in a sample is not evidence, and generated context must not contradict a rule the consumer already enforces. The system was itself built by an orchestrating agent with implementer and reviewer subagents; the failures on that side were scope failures. Limitations and the planned evaluation are stated.

Article: andrii.korkoshko.com. Technical report, version 1.0.