I've been building a Rails engine for managing LLM-powered agents in production. The main problem it solves: you define agents with a Ruby DSL, and everything else — cost tracking, retries, fallbacks, circuit breakers, caching, multi-tenancy, and observability — is handled by a middleware pipeline.

It ships with a mountable dashboard that shows execution history, spending charts (cost/tokens over time), per-agent stats, model breakdowns, and multi-tenant budget management with hard/soft enforcement.

Works with OpenAI, Anthropic, Google, ElevenLabs via RubyLLM. Supports text agents, embedders, TTS, transcription, image generation, message routing, and agent-as-tool composition.

v3.7, MIT licensed, ~4000 specs. Would appreciate feedback on the DSL design and middleware architecture