Agent Interaction Models

A conversational interface with a persistent sidebar showing background work. Ask questions naturally — "find cheaper alternatives to our top 10 ortho implants" — while the agent queries CMS datasets, checks vendor catalogs, and runs price comparisons in the rail. The chat is your foreground; the rail is the agent's workspace.

Data flows through stages: Upload → Parse → Enrich (CMS matching) → Analyze → Recommend → Act. Each column is an agent phase with its own memory and tools. Cards represent work items — a batch of uploaded invoices, a contract renewal, a vendor RFQ — moving through the pipeline as agents process them autonomously.

Jupyter-like cells where each cell is an agent action. Write SQL against multi-TB CMS/payer datasets, interleave with natural language instructions, and let the agent generate code to bridge the gaps. Some cells are "live" — continuously monitoring query results. Built for the ops analyst who thinks in queries but needs agent intelligence.

A vertical timeline of everything the agent has done, is doing, and plans to do. Every tool call, API request, data query, and decision is logged with full context. Critical for healthcare compliance — auditors can trace exactly how a procurement decision was made. Users can branch off at any point to explore alternatives.

Define what matters and let agents wake up when it happens. Price drops > 10% on a tracked implant? New CMS fee schedule published? Contract expires in 30 days? Patient volume spikes above threshold? Each trigger is a card with conditions, actions, and history. The agent sleeps until the world changes.

Two panes. Left: the agent's reasoning — tool calls, code execution, API requests, intermediate thoughts. Right: the clean output — formatted tables, recommendations, reports. Users can intervene on either side: correct the reasoning or edit the output. Transparency without clutter.

Multiple specialized agents working in parallel with shared memory. A Pricing Agent queries CMS data. A Compliance Agent checks regulatory requirements. A Vendor Agent manages RFQs. A Data Agent ingests uploads. They coordinate through a central memory hub, and the materials manager sees them all.

Each agent gets a persistent workspace tab that maintains state across sessions. "Knee Implant Analysis" has been running for 3 days, accumulating data. "Q4 Contract Renewals" is active now. "CMS Data Refresh" runs nightly in a background tab. Pin important workspaces, group related ones, and let agents work while you're away.

Minimal UI. The agent works silently — monitoring CMS feeds, tracking price changes, watching contract deadlines, scanning uploaded data for anomalies. It only surfaces when something needs human attention, ranked by urgency and business impact. Notification-first, action-oriented. For the manager who doesn't want another dashboard.

The report IS the interface. A document-centric model where agents build sales collateral, cost analyses, and procurement briefs as living documents. Users interact with sections directly — "deepen this analysis", "add a CBSA benchmark chart here", "rewrite for the CFO". The agent has full context of the document and multi-TB CMS datasets. Each section tracks its own data provenance, approval status, and revision history. Designed for the first use case: building sales collateral reports with custom, on-demand analyses embedded inline.

An async analysis fulfillment model. Users submit natural-language requests — "What's our TKA cost vs CBSA average?", "Generate vendor comparison for hip implants", "Show payer rate trends for Q4" — and they queue up for agent processing. Results are delivered as formatted, composable blocks: tables, charts, narratives. Completed briefings can be pulled into reports, shared, or used as evidence for contract negotiations. Like having a research analyst on staff, but one that can query 2TB of CMS data in minutes.

Every agent output passes through human review before it becomes official. Draft sections sit in a review queue with full data provenance — the user sees exactly what CMS data backed the numbers, which queries ran, and what assumptions were made. Approve, reject with feedback, or edit directly. Critical for sales collateral (you can't send wrong numbers to a prospect), procurement decisions (a bad RFQ costs real money), and compliance (auditors need the approval chain). Approval triggers downstream actions: approved report sections get assembled, approved RFQs get sent.