The author presents a branch‑only experiment that adds a fully isolated support_triage module to the OrKa orchestrator. The module registers seven custom agent types—envelope_validator, redaction, trust_boundary, permission_gate, output_verification, decision_recorder, and risk_level_extractor—without any modifications to the core runtime. By enforcing a structured input envelope, the workflow validates intent early, rejecting disallowed actions such as issue_refund and change_account_settings at the schema level.

PII redaction is performed deterministically, replacing email and phone data with placeholders and recording counts. The module also surfaces prompt‑injection attempts, though the current detection agent reports false negatives, highlighting a target for improvement. Parallel retrieval is demonstrated through a fork‑and‑join pattern that merges knowledge‑base search and account lookup results in a deterministic manner, providing clear observability of pending and completed agents.

Metrics for latency, token usage, and model cost are captured per node, enabling engineering decisions about scaling, caching, and local‑first execution. Decision recording and output verification ensure that each workflow step is replayable and auditable. Overall, the proof shows that OrKa can evolve via feature modules, preserving core stability while allowing rapid iteration on specialized agents.