FHIR-native prior authorization agent. Reads patient FHIR records, matches payer criteria, scores evidence, and drafts complete PA letters and appeals.
AuthBridge defines 9 tools with mostly complete schemas and descriptions. Tool names follow verb_noun convention (fetch_, lookup_, score_, draft_, verify_, generate_, run_). Descriptions are substantive (100-300 chars), explaining WHAT and WHEN. However, parameter descriptions are inconsistent, some params like 'patient_context' and 'pa_criteria' lack type hints in descriptions, forcing LLMs to infer structure. Output schemas are not documented in tool definitions. Error handling is minimal; no recovery guidance or actionable error messages visible. Security: no evidence of secret injection for API keys (GITHUB_TOKEN hardcoded in get_async_client). Composition is sound, tools chain logically (fetch → lookup → score → draft). Missing: per-tool output schema documentation, error classification, and confirmation patterns for destructive writes (draft_pa_letter, draft_appeal_letter, ingest_payer_policy).
Drafts a formal appeal letter rebuttal with guideline citations.
Drafts a justification letter with urgency headers and FHIR evidence trail.
Returns: Structured dict with patient_info, conditions, active_medications, medication_history, observations, procedures, allergies, fetch_errors.
Generates a plain-language PA status summary for the patient. No clinical jargon, no ICD codes. Designed for patient portal delivery.
Ingests unstructured payer policy text, extracts structured PA criteria using an LLM, and dynamically updates the payer_criteria.json database.
Looks up clinical PA requirements. Covers 16+ major therapeutic drugs.
No output schemas documented. Tool descriptions state WHAT is returned (e.g., 'Structured dict with patient_info, conditions...') but lack formal JSON Schema definitions. LLMs cannot plan downstream tool calls or extract fields reliably without documented output types.
Complex object parameters (patient_context, pa_criteria, match_result) lack type descriptions. Descriptions say 'Structured patient FHIR context' but do not specify required fields, nested structure, or format. LLMs cannot validate or construct these objects without explicit schema.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 62 | <=2025-11-25 | v2 |
Runs the complete AuthBridge prior authorization workflow in a single call. Fetches FHIR patient context, looks up PA criteria, scores clinical match, and drafts a complete PA letter.
Analyzes patient record against PA criteria using clinical reasoning. Includes CMS-0057-F urgency detection and FHIR evidence citations.
Audits a generated PA letter against the FHIR evidence trail. Flags any clinical claim that cannot be traced to a source FHIR resource. Implements AI self-verification to prevent hallucination before physician review.
No error handling or recovery guidance. Code uses @retry decorator but tool definitions do not document failure modes, retryability, or what to do if FHIR server is unreachable or PA criteria lookup fails. Agents cannot self-correct.
Destructive/write tools (draft_pa_letter, draft_appeal_letter, ingest_payer_policy) lack confirmation or dry-run patterns. No indication that these modify state or require user approval before execution. Agents may generate letters or ingest policies without safeguards.
GITHUB_TOKEN exposed in code (get_async_client). Credentials must never appear in tool parameters or implementation. Use server-side secret injection via environment variables or vault. Current pattern risks token leakage in logs and traces.