Authors: Camy Pearson, Michael Segner (Anthropic) URL: https://claude.com/resources/articles/how-curative-is-building-a-national-provider-network-with-agents-on-claude Published: November 10, 2026 Series: “How startups build with Claude” Product: Claude Platform, Claude Code
Curative, a health insurance startup with 1.275B valuation), built “Gwen” — a hierarchical multi-agent system on Claude that signs 2,000-2,400 provider contracts per week, up from 2,100 in all of 2025 (~50x faster).
Company Context
- Model: Health insurance with 0 deductibles for in-network care (after Baseline Visit)
- Origin: COVID-19 lab startup that processed ~10% of all US COVID tests; worked with nearly every insurer; pivoted to health insurance
- Problem: Building a national provider network requires signing 50,000+ contracts. At 2,100/year (double their prior year), it was a 30-year build.
“Sometimes naivety is good in startups, because you don’t know that it’s impossible. And then along the way you figure out a way to make it not impossible.” — Fred Turner, CEO
Gwen: The Provider Contracts Agent
Gwen handles a provider deal from initial outreach to contract signature, both inbound and outbound. Each run is isolated to a single provider — nothing from other deals influences it.
Workflow (Inbound Example)
- Practice fills out form on Curative’s website
- Gwen researches the practice
- Gwen proposes rates (using purchased claims data + price-transparency files)
- Gwen presents market rate data to aid provider negotiations
- If Gwen wants above-template rates → rate reviewer approves
- When agreed, Gwen sends contract for e-signature
Agent Hierarchy
| Agent | Model | Role |
|---|---|---|
| Gwen | Claude Opus | Lead deal agent — full provider lifecycle |
| Rate reviewer | Claude Sonnet | Approves any rates above standard template |
| Haiku judge | Claude Haiku | Additional quality check on outputs |
| Human | — | Escalation for non-standard pricing and edge cases |
Key Design Decision: Second Agent vs More Authority
“For a long time there was a big approval queue where Gwen had to get rates approved if she wanted to go a certain amount above the template, but we relaxed over time. Having a second agent review rates worked better than just giving Gwen more authority.” — Fred Turner
This is a critical pattern: specialist reviewer agents outperform a single high-authority agent for decisions with meaningful stakes.
Architecture Principles
- Context isolation per work unit: Each Gwen run is scoped to one provider. No cross-contamination between active negotiations.
- Tiered model usage: Opus for lead agent (quality), Sonnet for reviewer (cost/quality balance), Haiku for judges (cost efficiency)
- Human escalation preserved: Non-standard pricing escalates to humans. The system handles the volume; humans handle the exceptions.
- Stack: Claude Platform + Claude Code + bespoke claims system
Results
| Metric | Value |
|---|---|
| Contracts signed by Gwen (total) | 23,000 |
| New contracts per week | 2,000-2,400 |
| 2025 full-year contracts (pre-Gwen) | 2,100 |
| Speed improvement | ~50x |
Key Lessons for Multi-Agent Design
- Isolate context per work unit — running one deal per agent instance prevents cross-deal contamination
- Second-agent review beats giving the lead agent more authority — for consequential decisions (rates, commitments), a fresh reviewer is more reliable than a permissive single agent
- Model tiering is essential — use the highest-capability model only for the highest-value tasks; reviewers and judges can use smaller models
- Human escalation is a feature, not a failure — define clear escalation triggers and preserve them even as the system matures
- Contracts as code — structured, templated documents are well-suited to agent-based negotiation
Key Takeaways
- Gwen signs 2,000-2,400 provider contracts per week (was 2,100 for all of 2025)
- Hierarchical multi-agent: Opus lead + Sonnet reviewer + Haiku judge + human escalation
- Context isolation per provider is critical for accuracy in high-volume agent systems
- Multi-agent review outperforms single-agent high authority for consequential decisions
- Real-world production scale: not a demo — 23,000+ contracts signed
- Built on Claude Platform + Claude Code; bespoke claims system integration