A large health plan engaged ArtiSoft to design an agentic framework on Snowflake Coco that monitors and maintains its healthcare data pipelines — cognitive judgment handled by agents, deterministic execution and final authority kept with people.
Claims, clinical, eligibility, and reference feeds move through dozens of pipelines. Monitoring and maintaining them fell almost entirely to a small group of senior engineers.
On-call engineers triage hundreds of threshold alerts a week; most are benign variation.
Upstream feeds change shape without notice, breaking or quietly corrupting downstream transforms.
Every incident means correlating logs, lineage, and tickets across disconnected tools.
Pipeline fixes and new development compete for the same engineers, so both backlogs grow.
A canonical domain model is established before any agent is built, so agents emit structured, verifiable objects rather than plausible-sounding prose. Every process step is then classified as cognitive or deterministic — deterministic steps stay with conventional systems, cognitive steps go to agents, and every transition into execution requires human supervision. JetStore is the system of record for that boundary: it owns the domain model and executes every deterministic task, so agents are never asked to do a compute platform's job.
Every capability is bound to a tier at design time, enforced by which tools are present in an agent's catalog. Tiers advance only against measured evidence, and promotion is automatically reversible if outcomes turn adverse.
| Tier | Agent may | Human role |
|---|---|---|
| T0 — Observe | Read metadata and code; record analysis to dashboards | None; aggregate review only |
| T1 — Advise | Produce advisory hypotheses and incident narratives | Consumer of advice; final decision-maker |
| T2 — Propose | Generate staged, non-executing artifacts (diffs, plans) | Reviewer and author-of-record |
| T3 — Act with approval | Execute specific actions after authorization | Approver for every instance |
| T4 — Bounded autonomy | Execute reversible actions within an authorized envelope | Post-hoc reviewer; envelope owner |
Observer and Triage agents cut time spent correlating evidence during incidents. Schema Drift and Transformation Engineer agents turn maintenance from ground-up authoring into review. Reduced firefighting frees senior time for roadmap work.
Designed into the framework rather than left to prompt instructions. Each is enforced where it cannot be talked out of — in the tool catalogue, the platform roles, or the audit path.
PHI enters agent context only when essential, preferring de-identified or aggregate representations.
If a write to the audit store fails, the associated action does not happen.
On production transformation code; the approver may not be the engineer who initiated the task.
External feed content is treated as non-instructional, so no feed can invoke an agent capability.
A single control suspends all agent execution while the underlying platform keeps running.
Agents cannot alter their own prompts, permissions, or guardrails.
We'll walk through the domain model, the monitoring and maintenance agent designs, and a Phase 0 plan sized to your data ecosystem on Snowflake Coco.
This engagement was designed for the client's platform. ArtiSoft's own recommendation is JetStore Agentic AI, where the reasoning layer runs on inference you host and no external LLM provider is required — the platform accommodates either.
Email us to discuss the domain model, autonomy tiers, or a Phase 0 scope for your data platform.
info@artisoft.io