NoAction
No meaningful clinical work is required; the reason and evidence are still recorded.
Part 04BMaster architecture · Decide what matters now
Filter broadly. Reason selectively. Explain every priority.
Given an updated longitudinal patient state, Praxis determines whether meaningful work exists, which bounded specialists should reason, how urgency compares across the panel and what evidence supports the result.
DecisionCandidate
A versioned recommendation with urgency, evidence, confidence, safety requirements and expiration.
End-to-end control flow
Services control the work; agents perform bounded reasoning
01
Trigger & freeze
Start from a schedule, clinician request or meaningful PatientStateChanged event; freeze the patient and panel snapshot.
02
Gate & pre-rank
Apply safety overrides, eligibility rules and cheap longitudinal scoring before any agent runs.
03
Build context
Assemble a versioned, task-specific ContextPacket from clinical, evidence, workflow, preference and guideline memory.
04
Route specialists
Match clear clinical changes deterministically; use a small model only when unstructured intent is ambiguous.
05
Reason selectively
Run one to three relevant specialists in parallel against the same patient-state version.
06
Reconcile & ground
Merge compatible assessments, preserve conflicts and verify every clinical claim against current evidence.
07
Validate & rank
Apply deterministic safety and freshness policies, then produce an explainable DecisionCandidate or NoAction.
08
Hand off durable work
Temporal carries the validated candidate into approval and coordination.
Panel funnel · illustrative
Do not reason over all 1,200 patients at request time
1,200
deterministically scored
30
strong candidates
8
need fresh reasoning
Fresh DecisionCandidates are reused. Agents run only for high-ranking patients whose reasoning is missing, stale or conflicting.
Runtime agent count
No meaningful change0 agents
Single-condition case1 specialist + validator
Multi-condition case2–3 specialists + reconciler + validator
Unresolved complexityfrontier reconciliation + human
Context Builder
Agents start stateless—with the minimum trusted memory they need
01
Clinical memory
PostgreSQL
Current snapshot, personal baselines, longitudinal trends, active conditions and open care gaps.
02
Evidence memory
pgvector + text search
Notes, reports, messages and source-linked document chunks.
03
Workflow memory
PostgreSQL + Temporal
Previous recommendations, approvals, rejections, completed actions and current execution state.
04
Preference memory
PostgreSQL
Explicit clinician communication, routing and workflow preferences—separate from clinical truth.
05
Knowledge memory
Versioned knowledge store
Approved clinical guidelines and organization policies with effective dates and versions.
Versioned ContextPacket
A frozen, agent-specific view—not the entire chart
context_idpatient_idworkflow_idtriggeras_ofstate_versionrelevant_changelongitudinal_windowevidence_citationsguideline_versionsprior_outcomespreferencesmissing_or_conflicting_dataallowed_toolsrisk_leveltoken_budgetexpires_at
Agents fetch additional facts through typed tools. They never directly modify clinical, evidence, workflow or preference memory.
Agent runtime
Ten registered roles; two to five usually execute
IDAgentTypeResponsibility
01
Clinical Coordinator
Optional
Scopes complex or ambiguous multi-condition reasoning.
02
Diabetes & Metabolic
Specialist
Reviews metabolic trends, care gaps and related complications.
03
Cardiovascular
Specialist
Reviews cardiovascular changes and relevant longitudinal risk.
04
Kidney
Specialist
Reviews renal trends, gaps and cross-condition implications.
05
Medication Safety
Specialist
Reviews medication changes, monitoring and interactions.
06
Preventive Care
Specialist
Reviews screening, vaccination and preventive gaps.
07
Reconciler
Shared
Combines compatible assessments and preserves unresolved conflicts.
08
Grounding Reviewer
Shared
Verifies claims, citations, freshness and guideline support.
09
Inbox Drafting
Workflow
Converts a validated decision into a reviewable response draft.
10
Practice Improvement
Workflow
Finds practice-level patterns from governed aggregate evidence.
Agent Directory
Approved, versioned manifests
agent_id + version
supported workflows
condition + event triggers
required context
input + output schemas
allowed tool IDs
model policy
risk class
token + timeout budget
prompt version
evaluation suite
lifecycle status
Draft
Offline eval
Shadow
Canary
Active
Disabled
Tool Gateway
Typed access—not arbitrary database queries
Patient state
Readsnapshot · timeline · labs · medications · care gaps · prior outcomes
Evidence
Readdocument search · message thread · source chunk · citation resolver
Knowledge
Readguideline search · approved section · organization policy
Calculations
Computeapproved measure · personal baseline · trend · interaction check
Preparation
Preparemessage draft · task draft · order request · EHR note draft
Execution
Approval onlysend message · create task · submit order · write note
Controlled delegation
Need more data?
Call a scoped read tool.
Need another specialty?
Return a SpecialistRequest.
Router decides
Check relevance, duplication, risk and budget before Temporal starts it.
No direct agent spawning · no recursive delegation · maximum depth one · maximum three specialists
Model routing + prioritization
Complexity chooses the model; risk chooses the safeguards
Tier 0
No model
Rules, calculations, safety gates and ranking
Deduplicate an event, compare to baseline and score the panel.
0 tokens
Tier 1
Small model
Classification and extraction
Extract symptom, timing and likely specialty from an ambiguous message.
≤ 2k in · 150 out
Tier 2
Mid-tier model
Normal specialist reasoning, drafting and grounding
Assess one condition using relevant trends, evidence and guidelines.
≤ 8–12k in · 1.2k out
Tier 3
Frontier model
Complex multi-condition reconciliation
Resolve conflicting specialists across a long, contradictory timeline.
≤ 24k in · 1.5k out
Explainable priority · initial hypothesis
0–100 factor score
Potential clinical severity30%
Change from personal baseline25%
Time sensitivity20%
Unresolved work15%
Actionability10%
Safety overrides run first. Confidence and freshness remain separate; uncertainty never silently lowers a serious case.
Freshness + concurrency
Never publish against stale state
Version dependencies
Candidate records input state, trigger events, guideline and agent versions.
Dependency invalidation
Only related labs, medications, messages or completed actions invalidate it.
Signal active workflow
Temporal groups related events; critical events bypass batching.
Compare before activation
A changed relevant state version rejects publication and reruns affected steps.
DecisionCandidate contract
priority tier + score breakdown
what changed
why it matters now
recommended next step
evidence + citations
confidence + missing data
required approval
state, agent, model + guideline versions
created_at + expires_at
Worked escalation trace
One event moves through every control layer
Illustrative scenario: a new laboratory result arrives and the patient reports dizziness after a recent medication change.
01Ingestion
Two new signals
A new lab result arrives and the patient reports dizziness after a medication change.
02Tier 0
Deterministic gate
Verify identity, update longitudinal state, compare to baseline and apply critical-event policy.
03Tier 1
Classify ambiguity
Extract symptom, timing and the likely kidney + medication-safety specialties from the message.
04Context
Build shared context
Load renal trend, medication timeline, prior actions, approved guidance and exact citations.
05Tier 2
Reason in parallel
Kidney and Medication Safety Agents independently produce typed, cited assessments.
06Tier 3
Escalate conflict
If assessments materially disagree, frontier reconciliation resolves what it can and preserves uncertainty.
07Human gate
Validate and approve
Ground every claim, apply safety policy and present evidence, conflicts and missing data to the clinician.
08
Clinician chooses
Approve, edit, reject, request more information or escalate. No action executes before the required gate.
Reliability + evaluation
Fail visibly, replay safely and test every boundary
Failure behavior
Transient tool failure
Retry with backoff and the same idempotency key.
Invalid model output
One schema repair; then approved same-tier fallback.
Required specialist fails
Stop completion and route to human review.
Context becomes stale
Reject activation and rerun only the affected steps.
Grounding fails
One revision cycle; then human escalation.
Worker crashes
Temporal resumes from recorded workflow state.
Evaluation layers
Deterministic unit tests
routing · scores · freshness · policies · idempotency
Tool contract tests
schemas · scope · permissions · timeouts · citations
Agent evaluation suites
correctness · abstention · grounding · calibration · cost
Multi-agent tests
agreement · conflicts · partial failure · unsupported synthesis
Workflow replay tests
crashes · duplicate events · cancellation · late approval
Deployment gates
offline → replay → shadow → clinician review → canary
Versioned assumptions
Values are hypotheses until evaluations calibrate them
Priority-score weights and tier boundaries
Critical-event and approval policies
Panel funnel sizes and maximum three specialists
Candidate expiration and invalidation rules
Model assignment, token budgets and escalation thresholds
Latency, cost and confidence targets
Decision principle
Tools fetch facts. Agents reason within boundaries. The router delegates. Independent reviewers validate. Humans authorize consequential action.