How Curative is building a national provider network with agents on Claude

Curative deployed an agent hierarchy built on Anthropic's Claude to accelerate health plan provider contract signings to 2,000 to 2,400 agreements per week, according to claude.com's report. The employer health insurance startup, which raised a $150 million Series B at a $1.275 billion valuation, previously signed only 2,100 provider contracts during all of 2025. Across its rollout, the autonomous pipeline has executed 23,000 provider agreements, signing contracts about 50 times faster than the company's manual pace.

This report examines how Curative structured its multi-model architecture, automated legal document redlining, and rebuilt proprietary insurance backends for engineering leaders and healthcare operators. It details the operational guardrails, token expenditures, and human-agent workflows required to construct a national network of healthcare providers.

01Tiered model roles for contract negotiations

Each run of Curative's primary contracting agent, named Gwen, handles a single healthcare provider deal powered primarily by Claude Opus. When a medical practice submits an inquiry, Gwen analyzes commercial claims datasets and mandatory public price-transparency files to propose competitive reimbursement rates. Operating in isolated per-deal runs prevents cross-contamination between competing medical practices during price negotiations.

Curative surrounds Gwen with secondary models to enforce governance, including a Claude Sonnet rate reviewer that clears two-thirds of pricing exceptions and a Claude Haiku judge that inspects outgoing drafts. Finalizing an individual contract consumes approximately 20 million tokens, representing a 90 percent drop from early operational costs. As Fred Turner explained, having a second agent review rates worked better than just giving Gwen more authority.

02Staff productivity and broker distribution workflows

Curative expanded its internal contracting personnel from 45 to roughly 55 employees while lifting individual worker productivity by 2.5 times. Human negotiators delegate cold outreach and administrative paperwork to Gwen while focusing personal attention on telephone negotiations with warm provider leads. The company aligns compensation by paying human contracting staff performance incentives on deals closed by autonomous agents.

Employer sales expand through an independent Opus-based agent called Brad, which coordinates with health insurance brokers to manage policy quotes. Brad accelerates underwriting by extracting employee census data directly from raw spreadsheets, eliminating manual data cleaning formerly handled by human staff. Curative also deployed automated dispute handlers that resolve emailed or faxed provider billing disputes in under 30 minutes, down from 90 days.

03Core claims architecture on cloud infrastructure

A team of 13 software engineers rebuilt Curative's proprietary claims engine, prior authorization infrastructure, and financial accounting backends using Claude Code. Traditional enterprise estimates had originally sized developing a bespoke claims platform at 100 engineer-years of manual effort. The completed architecture runs entirely in Python hosted on Amazon Web Services, exposing native application programming interfaces and Model Context Protocol endpoints.

Rather than forcing autonomous agents into rigid legacy platforms, Curative engineered its claims backends so generative models interact natively with database services. Fred Turner noted that the claim system is Python running in AWS with an API and Claude wrote it, so it understands it. This unified architecture enables rapid deployment of new insurance workflows without expanding foundational engineering headcount.

04Fax communication channels and execution safeguards

Curative simplified agent tool architectures by eliminating sprawling specialized toolsets in favor of a single general Python execution environment. Fred Turner observed that redlining is quite complex, finding that models work best when restricted to running Python to execute document edits with a docx library. This single-tool design lowered cognitive overhead for the model and prevented stalled multi-step execution chains.

Integrating physical facsimile transmission tools more than doubled Gwen's contract closure rates across medical practices. Chief technology officer Isaac Turner explained that while practices maintain email, providers are genuinely comfortable working through fax. Curative also established isolated execution boundaries, ensuring each run of Gwen retains zero conversational memory across distinct provider negotiations to prevent compounding factual errors.

05Protected health information and autonomous initiatives

Curative isolated agents from protected health information during early development phases to refine decision algorithms within low-risk administrative envelopes. As workflows expanded into clinical domains, backend controls filtered sensitive records while comprehensive auditing mechanisms logged every automated action. The company negotiated formal Business Associate Agreements under HIPAA requirements, leveraging Anthropic's compliant cloud infrastructure to safeguard member privacy across healthcare integrations.

Unprompted system behaviors emerged as models solved operational obstacles without explicit rule programming. When prospective clinic partners requested internal informational pamphlets, Gwen utilized its Python script runner to generate custom distribution flyers independently. In another instance, after an in-network provider office improperly turned away an enrolled patient, Gwen contacted clinic administrators to request staff retraining and later verified member admission.