Generative LLMs excel at syntax and typical imperative programming idioms. However, when building resilient concurrent backends with Elixir and OTP, LLMs frequently introduce subtle architectural flaws.

Common LLM Failures in OTP Systems

1. Treating GenServers as State Machines for Everything

LLMs frequently propose a single monolithic GenServer to manage business logic. In reality:

  • GenServers should model concurrency bottlenecks or state that outlives a request, not standard functional domain logic.
  • Pure Elixir modules (functional cores) should handle business logic; GenServers should act solely as thin process coordinators (imperative shells).

2. Violating OTP Supervision Hierarchies

When an unhandled crash occurs, OTP restarts the worker process according to its supervision strategy (:one_for_all, :one_for_one). AI models frequently wrap GenServer calls in catch-all try/rescue blocks, blinding the supervisor to catastrophic failures and causing zombie worker states.

The Resilient Pattern: Functional Core, Process Shell

By isolating pure functions and relying on “Let it Crash” principles with Oban or OTP supervisors, systems achieve self-healing properties that imperative code cannot match.