Start with panel prioritization because it forces the hardest decisions—incrementality, ranking, cost, safety, freshness and trust—while reusing today’s single-patient reasoning core.
Delivery horizons
Two weeks
A clinician can review a safe weekly queue generated from synthetic/fixture panels.
Scheduled Temporal parent workflow
SQL/change pre-filter + bounded top-candidate fan-out
Reuse existing single-patient reasoning core
Evidence-backed ranked queue; human approval only
Audit trail, budget ceiling and engineered fixture panel
Two months
Panel and inbox workflows share durable state, event triggers and deterministic evaluation.
PatientStateChanged invalidation + freshness policies
Inbox classify/draft/hold/escalate workflow
Typed agent/tool contracts + improved grounding reviewer
Feedback and explicit preference/constraint records
Replay, shadow, canary and cost/latency dashboards
Two quarters
Three workflow shapes run on one governed, measurable backplane.
Practice-improvement analytics and patient drill-down
Calibrated prioritization by organization and population
Selective low-risk operational autonomy
Outcome-linked memory with review and expiry
Broader connectors, policy layering and continuous evaluation
First proof
Can the system find valuable work without flooding the clinician?
Second proof
Can event-driven panel and inbox work share contracts, state and evaluations?
Third proof
Can safe feedback improve prioritization and practice insight over time?
Bonus · Shared backplane migration
PatientState, ContextPacket, DecisionCandidate, SpecialistRequest, ActionRequest and evaluation events.
Context Builder, Tool Gateway, Agent Directory, reconciler, grounding and action policy become host-agnostic modules.
Current synchronous session runtime and Temporal async runtime call the same versioned core.
Run the new path beside the current batch path; compare actions, citations, cost and failures.
Panel first, then inbox, then practice insights; no rewrite and no team freeze.
Synchronous runtime
One clinician turn; low-latency response and approval
Shared reasoning core
Context · tools · agents · reconcile · ground · policy · eval hooks
Async Temporal runtime
Many durable units across patients, channels and time—calling the same core
Deliberate technical debt
Hand-tuned ranking weights
Fast, explainable starting point
Calibrate from adjudicated panels; retain deterministic guardrails
PostgreSQL + pgvector
Simpler security and transactional provenance
Split only when measured recall/latency/scale thresholds fail
Limited condition roster
Ship safe depth for common chronic disease
Add agents only with owner, tools, eval set and lifecycle plan
Human approval for every clinical action
Controls blast radius while quality is learned
Expand only narrow, reversible operational autonomy
Partial event coverage
Start with labs, medications, messages and scheduled review
Add claims/transitions after source-specific quality contracts
Explicit non-goals
Do not replace the EHR or authoritative source systems.
Do not run frontier reasoning across every patient and condition.
Do not create a separate “Praxis Async” reasoning stack.
Do not learn clinical truth from rejection behavior alone.
Do not auto-send clinical messages or execute high-risk actions.
Do not build a generic agent marketplace or recursive agent hierarchy.
Do not claim clinical thresholds without clinician ownership and evaluation.
Do not introduce a dedicated vector database before pgvector proves insufficient.
Assumptions held constant
PostgreSQL, object storage, Temporal, hosted observability and frontier APIs already exist.
EHR write-back supports only a narrow subset of action types.
Every clinical action has a named accountable human.
No PHI reaches a vendor without a BAA; logs and traces minimize PHI.
Clinician attention—not compute—is the scarcest resource.
Critical clinical thresholds, labels and autonomy policies require clinician ownership.
All patients and examples used in development and evaluation are synthetic or approved fixtures.
Final position
Understand
Versioned patient state
Decide
Selective, grounded reasoning
Coordinate
Durable human-owned work
Improve
Measured feedback with guardrails
Recommendation
Ship a safe, explainable weekly panel queue first. Use what it teaches about state, cost, trust and evaluation to earn inbox automation and practice-level learning.
Working-session stance
These boundaries are firm; the values are hypotheses. I would defend the safety and state model, then gladly redesign thresholds, topology and sequencing as Altitude changes the constraints.