Panel request example
1,200
rule-scored
30
candidates
8
fresh reasoning
Presenter view
Decision architecture speaker notes
Explain how Praxis narrows a large panel, gives stateless agents trusted context, escalates model effort only when necessary and keeps consequential action behind a clinician.
Opening cue
“The previous layer gave Praxis a trustworthy longitudinal patient state. This layer decides whether that change deserves attention, which specialists should reason about it and how it should compare with work across the rest of the panel.”
01 · Decision boundary + panel funnel
Do cheap work broadly; expensive reasoning selectively
“The input is an updated longitudinal patient state. The output is either NoAction or a versioned DecisionCandidate. A candidate is not an executed clinical action—it is an explainable proposal that can enter a durable approval workflow.”
How to walk the control flow
- 01
Freeze the panel and patient-state versions.
- 02
Run safety overrides, eligibility rules and deterministic scores.
- 03
Reuse fresh candidates; identify missing, stale or conflicting reasoning.
- 04
Build context, route specialists, reconcile, ground and rank.
Key point: a clinician request does not create 1,200 agent sessions. Most patients stop after deterministic computation, and already-fresh candidates are reused.
02 · Context Builder
Longitudinal state becomes task-specific memory
Clinical memory
PostgreSQLCurrent snapshot, personal baselines, longitudinal trends, active conditions and open care gaps.
Evidence memory
pgvector + text searchNotes, reports, messages and source-linked document chunks.
Workflow memory
PostgreSQL + TemporalPrevious recommendations, approvals, rejections, completed actions and current execution state.
Preference memory
PostgreSQLExplicit clinician communication, routing and workflow preferences—separate from clinical truth.
Knowledge memory
Versioned knowledge storeApproved clinical guidelines and organization policies with effective dates and versions.
“Agents are stateless. The Context Builder gives each one a frozen, versioned packet containing only the relevant slice of patient history, evidence, prior workflow outcomes, explicit preferences and approved knowledge.”
context_idpatient_idworkflow_idtriggeras_ofstate_versionrelevant_changelongitudinal_windowevidence_citationsguideline_versionsprior_outcomespreferencesmissing_or_conflicting_dataallowed_toolsrisk_leveltoken_budgetexpires_at
The longitudinal record stores baseline, trends, episodes, prior recommendations and outcomes. Temporal stores workflow progress. Neither is hidden inside an agent conversation.
03 · Agents, directory + tools
Logical roles in one governed runtime
Agent roster
Ten registered roles; usually two to five run.
01
Clinical Coordinator
OptionalScopes complex or ambiguous multi-condition reasoning.
02
Diabetes & Metabolic
SpecialistReviews metabolic trends, care gaps and related complications.
03
Cardiovascular
SpecialistReviews cardiovascular changes and relevant longitudinal risk.
04
Kidney
SpecialistReviews renal trends, gaps and cross-condition implications.
05
Medication Safety
SpecialistReviews medication changes, monitoring and interactions.
06
Preventive Care
SpecialistReviews screening, vaccination and preventive gaps.
07
Reconciler
SharedCombines compatible assessments and preserves unresolved conflicts.
08
Grounding Reviewer
SharedVerifies claims, citations, freshness and guideline support.
09
Inbox Drafting
WorkflowConverts a validated decision into a reviewable response draft.
10
Practice Improvement
WorkflowFinds practice-level patterns from governed aggregate evidence.
Agent Directory
“This is a registry of approved, versioned agent manifests: supported triggers, context schema, allowed tools, model policy, risk class, budgets, prompt version and evaluation suite. The router selects from this registry; it does not invent agents.”
Draft
Eval
Shadow
Canary
Active
Disabled
Tool Gateway
Each agent sees only four to eight typed tools needed for its task. Tenant and patient scope come from the workflow—not model arguments. There is no arbitrary SQL.
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
Simple delegation example
“The Kidney Agent can call a scoped tool for a missing renal trend. If it believes Medication Safety is needed, it returns a typed SpecialistRequest. The router checks relevance, duplication, risk and budget, then Temporal starts that specialist. The agent never directly creates another agent.”
Maximum depth one · maximum three specialists · no recursive spawning
04 · Model routing + ranking
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
Initial priority hypothesis
Explainable 0–100 score
Potential clinical severity30%
Change from personal baseline25%
Time sensitivity20%
Unresolved work15%
Actionability10%
“Safety overrides run before ranking. Confidence and freshness stay separate from priority, so missing information cannot quietly push a serious case down the list. Before activation, I compare the candidate’s dependency versions with the latest patient state.”
Freshness rule
A new relevant lab, medication change, message or completed action invalidates the affected candidate. Temporal reruns only the dependent steps; unrelated events do not discard useful work.
05 · Worked escalation + human loop
New lab plus dizziness after a 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 data or escalate.
“The lab is handled deterministically first. A small model extracts symptom and timing from the message. Kidney and Medication Safety reason independently. If they materially disagree, the Reconciler may use the frontier tier. The Grounding Reviewer then checks every claim and citation before the clinician sees it.”
Conflict behavior
Typed specialist outputs make overlap and disagreement visible. A deterministic pre-merge combines compatible claims. One reconciliation and grounding revision is allowed; unresolved material conflict becomes a human-review reason, not a fabricated consensus.
06 · Reliability, evaluation + assumptions
Every failure becomes an explicit state
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 stack
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
Say this explicitly
“The weights, thresholds, funnel sizes, freshness windows, agent limits, model tiers and token budgets shown here are starting hypotheses—not clinical truth. I would version and tune them through offline evaluations, historical replay, shadow mode, clinician review and canary rollout.”
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
Closing transition
“At this point Praxis has decided what deserves attention and produced an evidence-backed, versioned candidate. The next architecture question is how Temporal coordinates approval, waiting, retries and execution across hours or days.”