Agentic outbreak triage using LangGraph -- by Ankur Sharma, PhD
A 4-node LangGraph agent that triages infectious disease outbreak cases -- combining genomic analysis, contact linkage, risk scoring, and a self-correcting critic loop that re-evaluates when outputs are inconsistent.
Built around the April 2026 MV Hondius / Andes virus event -- the first confirmed human-to-human hantavirus transmission on a cruise ship.
Every run generates two outputs automatically -- no API key required:
Left: risk scores across all 4 outbreak scenarios, colour-coded by tier. Centre: event timeline from Patagonia exposure to MoH notification, showing the 48-72 hour human analyst window vs the under-2-hour agent window. Bottom: day-by-day contact cluster growth on MV Hondius -- index case expanding through confirmed contacts to at-risk HVAC zone occupants.
A structured PDF triage report is saved alongside the PNG:
reports/
ANDV-2026-001-risk-dashboard-2026-05-09.png <- dashboard chart
ANDV-2026-001-triage-report-2026-05-09.pdf <- full triage report
The PDF contains: risk tier banner, recommended action, genomic profile table, epidemiological linkage table, critic audit log, and interpretation notes -- all auto-populated from the agent state. Readable on mobile, laptop, or large display.
Standard outbreak surveillance tools are databases with search interfaces. You query them, they return data, a human decides what to do next. That loop takes 48–72 hours.
outbreak-agent is different. You give it a case and it gives you a risk-stratified
action, an audit trail, and a printed report. No human in the middle.
The MV Hondius scenario is exactly what it was designed for: Andes virus, the only hantavirus with confirmed human-to-human aerosol transmission, on a vessel with recirculated cabin air, rotating contact networks, and a 45-day incubation window during which infected passengers fly home globally. The contact list is not 3 named individuals -- it is every passenger sharing HVAC zones.
Human analyst: 48–72 hours to produce a risk-stratified contact list.
outbreak-agent: under 2 seconds. Deterministic. Auditable. Free to run.
START
|
v
+-----------------+
| genomic_node | Identifies clade & mutations from sequence or location heuristics
+--------+--------+
|
v
+-----------------+
| linkage_node | Resolves contact cluster, infers transmission mode
+--------+--------+
|
v
+-----------------+
| risk_node | Composite score 0-100, tier: LOW / MEDIUM / HIGH / CRITICAL
+--------+--------+
|
v
+-----------------+
| critic_node | Audits consistency, approved or loops back (max 3 iterations)
+--------+--------+
|
approved? --YES--> END (final_report + PNG dashboard + PDF report)
|
NO (critic_flags raised)
|
+-----------> genomic_node (re-evaluate with updated context)
You feed it the facts: patient, age, exposure location, ship, cabin, contacts, symptom onset, lab values, genome sequence if available. The agent runs 4 steps:
- Step 1 — Identify the virus.
genomic_nodedetermines the viral clade, flags mutations of concern, and assesses genome completeness. For MV Hondius: Andes virus, S-clade 2026, G2 glycoprotein shift (structural basis for aerosol transmission). - Step 2 — Map the contacts.
linkage_noderesolves the full contact cluster -- inferred from vessel layout, HVAC zones, and excursion groups -- and determines transmission mode. - Step 3 — Score the risk.
risk_nodeproduces a composite score (0–100) and assigns a tier: LOW / MEDIUM / HIGH / CRITICAL. For MV Hondius: 98/100, CRITICAL. - Step 4 — Audit the output.
critic_nodechecks the entire output for internal contradictions before anything reaches a decision-maker. Inconsistent outputs loop back for re-evaluation -- up to 3 times. Only a clean, consistent output is approved.
One shared state object (OutbreakState) flows through all 4 nodes. Each node
reads what it needs and writes back only its own outputs.
LangGraph vs LLM: LangGraph provides the graph engine -- state management, node
wiring, and the conditional edge that routes the critic's flag back to genomic_node
for re-evaluation. The 4 nodes are rule-based and deterministic; that is what makes
33 tests run free, fast, and offline. The architecture is designed so replacing any
rule-based node with a real LLM call requires changing one function -- not rewiring
the graph. Deterministic and auditable today, upgradeable without structural changes
tomorrow.
The critic_node is what makes this agentic rather than just a pipeline.
It enforces four consistency rules:
- ANDV + aerosol + LOW/MEDIUM tier -- flag: likely under-scoring
- Genome completeness < 70% with confident clade -- flag: re-sequence needed
- Cluster size > 8 with no exposure anchor -- flag: linkage unreliable
- UNKNOWN clade + CRITICAL tier -- flag: requires phylogenetic review
When any flag fires, the graph loops back for re-evaluation -- up to 3 times. This self-correction loop is the difference between an agent and a script.
git clone https://github.com/ankurgenomics/outbreak-agent
cd outbreak-agent
pip install -r requirements.txt
# Run the MV Hondius case -- generates PNG + PDF in reports/
python demo.py --case hondius
# Run all 4 mock cases
python demo.py
# Run free test suite (no API key needed)
pytest tests/test_nodes.py tests/test_graph.py -vTerminal output for the MV Hondius case:
Case: ANDV-2026-001
Clade : ANDV-S-clade-2026
Mutations : N-end-truncation-14aa, G2-glycoprotein-shift
Genome quality : 87.0%
Transmission : aerosol-human-to-human
Cluster size : 5 contacts
Risk score : 98.0/100
Risk tier : CRITICAL
Action : Immediate isolation. Notify MoH within 2 hours.
Activate IPC team. Contact trace all vessel passengers.
Approved by critic (loops: 1)
Dashboard : reports/ANDV-2026-001-risk-dashboard-2026-05-09.png
PDF Report: reports/ANDV-2026-001-triage-report-2026-05-09.pdf
| Case | Virus | Setting | Risk Tier |
|---|---|---|---|
hondius |
ANDV (Andes) | MV Hondius cruise ship, Apr 2026 | CRITICAL (98/100) |
high |
ANDV (Andes) | Patagonia family cluster | HIGH (85/100) |
low |
SNV (Sin Nombre) | Rural New Mexico | LOW (22/100) |
incomplete |
Unknown | Seoul (degraded genome) | MEDIUM -- critic flagged |
All mock cases run offline. No API key. No cost. Deterministic output.
Three-layer test strategy -- two layers are completely free:
| Layer | Command | What it tests | Cost |
|---|---|---|---|
| 1 | pytest tests/test_nodes.py -v |
Each node in isolation (23 tests) | Free |
| 2 | pytest tests/test_graph.py -v |
Full graph, all 4 mock cases (10 tests) | Free |
| 3 | pytest tests/test_smoke.py -v -s --smoke |
Live model API end-to-end | ~$0.01 |
# Free tests -- run these in CI
pytest tests/test_nodes.py tests/test_graph.py -v
# Expected: 33 passed
# Live smoke test (optional, needs API key)
export OPENAI_API_KEY=sk-...
pytest tests/test_smoke.py -v -s --smokeoutbreak-agent/
models.py -- OutbreakState TypedDict (shared state schema)
nodes.py -- 4 node functions: genomic, linkage, risk, critic
agent.py -- LangGraph StateGraph wiring + conditional critic edge
mock_data.py -- 4 pre-built outbreak scenarios
demo.py -- CLI runner (generates PNG + PDF automatically)
report_node.py -- matplotlib dashboard + ReportLab PDF generator
requirements.txt
tests/
test_nodes.py -- 23 unit tests (nodes in isolation)
test_graph.py -- 10 integration tests (full graph)
test_smoke.py -- live model smoke tests (optional)
reports/ -- auto-generated output directory
Agentic AI does not replace epidemiologists. It compresses the time between symptom onset and risk-stratified action -- from 48-72 hours to under 2 hours.
That compression matters because ANDV has a 45-day incubation window. Every hour of delay is more international passengers boarding flights home. The agent needs to be fast and auditable -- not perfect. The critic loop provides the audit trail. The PDF report gives the public health officer something to act on immediately.
The same architecture applies to any scenario where:
- Data arrives from multiple sources with inconsistencies
- A quality gate is required before a decision is made
- The decision must be explainable and documented
Outbreak surveillance is one domain. Clinical variant classification, fraud triage, and insurance risk scoring follow the same pattern.
Full write-up with biology, architecture, and design decisions:
When an AI Agent Boards a Cruise Ship: Hantavirus, LangGraph, and the Future of Outbreak Triage
Ankur Sharma, PhD -- Computational biologist and agentic AI engineer, Singapore.
Builds systems that connect genomic data to real-world decisions: outbreak surveillance, precision diagnostics, and autonomous research pipelines.
- LinkedIn: https://linkedin.com/in/ankurit
- GitHub: https://github.com/ankurgenomics
- Portfolio: https://ankurgenomics.github.io/agentic-genomics/
- Email: ankurs103@gmail.com
Copyright 2026 Ankur Sharma, PhD.
Code: Apache License 2.0
Blog posts and infographics: CC BY-NC 4.0 -- free to share with attribution, not for commercial use without permission.