Testing Agents Without a Model

A router seam behind one interface: a live LLM classifier in production and a deterministic classifier in CI, both feeding the same orchestration, gateway, HITL, and checkpoint stack

Testing Agents Without a Model A router seam behind one interface: a live LLM classifier in production and a deterministic classifier in CI, both feeding the same orchestration, gateway, HITL, and checkpoint stack Swappable implementations Shared runtime (identical in prod and CI) LLM Classifier · live model · Swappable implementations · production LLM Classifier live model production Deterministic Classifier · fixed rules · Swappable implementations · CI / offline Deterministic Classifier fixed rules CI / offline Router Interface · one seam · Architecture component · swappable Router Interface one seam swappable Orchestrator · routes turns · Shared runtime (identical in prod and CI) Orchestrator routes turns Policy Gateway · tool policy · Shared runtime (identical in prod and CI) Policy Gateway tool policy HITL Gate · human approval · Shared runtime (identical in prod and CI) HITL Gate human approval Checkpoint Store · resumable state · Shared runtime (identical in prod and CI) Checkpoint Store resumable state classify Legend Cloud Backend Security External Database

The Seam

  • • One router interface, two implementations
  • • Production binds the live LLM classifier
  • • CI binds a deterministic classifier

Same Stack Below

  • • Orchestration, gateway, HITL, checkpoints unchanged
  • • Tests exercise real control flow
  • • No model, no flakiness, no cost

Architecture diagram • Built with Archify • Create yours ↗ • Hover to trace • R route • Click to focus • +/− zoom • M radar • T theme • E export