Blood Panel Triage Team

Blood Panel Triage Team

Patient-education triage for blood lab results. The orchestrator runs a deterministic safety gate against the standard hospital 12-marker panic-value table, then either emits an emergency message or fans three cross-lab sub-agents out in parallel and integrates their outputs into a single patient-facing report.

Not a diagnostic tool. Not a replacement for a clinician.

Architecture

start (raw lab values)
   │
   ▼
triageOrchAgent (Zhipu GLM-5.1)
   │  ← check_critical_values Python tool (deterministic panic-value gate)
   │
   ├── if critical → EMERGENCY OUTPUT, stop
   │
   └── if not critical, call_sub_agent (parallel):
         ├── interpreterAgent   (Alibaba qwen3-max-thinking + harness_reasoning)
         ├── doctorpushAgent    (Anthropic claude-opus-4    + harness_anti_deception)
         └── differentialAgent  (DeepSeek deepseek-r1, no tools)

Roles

• triageOrchAgent — orchestrator + safety gate. Runs the check_critical_values inline Python tool against 12 markers (glucose, potassium, sodium, hemoglobin, platelets, WBC, INR, troponin, creatinine, lactate, calcium, magnesium). Critical → stops with an emergency message. Otherwise fans out the three sub-agents in parallel and pastes their replies verbatim into the final report.
• interpreterAgent — plain-language explainer. Locked to harness_reasoning (Ejentum cognitive harness) plus PubMed/Europe PMC literature search. Explains what each abnormal marker reflects in the body. Does not diagnose.
• doctorpushAgent — blunt second-opinion voice. Locked to harness_anti_deception plus NIH Clinical Tables lookups (conditions + LOINC lab tests). Refuses false reassurance — writes specific questions the patient should push their doctor on, including what the panel did NOT cover.
• differentialAgent — differential enumerator. No tools, no harness. Lists 3-5 conditions consistent with the abnormal pattern, each with a confirm/rule-out trio.

Setup

1. Get an Ejentum API key at ejentum.com and set it in each agent's MCP env field.
2. Connect LLM credentials in all four agent nodes (Anthropic, Alibaba/OpenRouter, Zhipu, DeepSeek).
3. Paste a CBC, CMP, or other blood panel as free text or JSON into the start node and run.

Example input

Hemoglobin 8.5, glucose 280, potassium 4.2, creatinine 1.8

Cross-lab design

Four different model providers reduce shared-RLHF blind spots: orchestration (Zhipu), interpretation (Alibaba), adversarial probing (Anthropic), and differential enumeration (DeepSeek).

Safety notes

• The panic-value table uses standard US hospital callback thresholds; individual labs vary.
• Reference ranges are adult, non-pregnant defaults.
• This workflow is patient-education software, not a diagnostic system. Discuss any abnormal value with a licensed healthcare provider before acting on it.

Built with heym + ejentum-mcp · agent-teams repo · ejentum.com