Cresta's Conductor Roughly Halves Agent Deployment Time by Packaging CX Expertise on the Claude Agent SDK

Cresta says its new agent builder, Conductor, roughly halved initial deployment time in early implementations — and the more useful detail is what it is built from. Anthropic’s Claude Blog profiles the customer experience AI company, founded in 2017 by Ping Wu, which has raised more than $270M, reports over $100M in ARR, and serves enterprises including United Airlines, CVS Health and Marriott. Conductor is a meta-agent: it guides a team from initial blueprint through implementation, evaluation and optimization of an agent, and it runs on the Claude Agent SDK, which Cresta uses as a general-purpose execution harness for context gathering, tool calling and code execution.

The pattern matters more than the vendor. Cresta did not build an execution engine. It layered its customer experience workflows, specialized tools and domain expertise on top of the SDK, then wrapped that in a builder that helps developers decide which workflows need deterministic control and which can stay flexible. Proven patterns get captured as reusable skills, so the second agent starts further along than the first.

The evaluation discipline is the part to copy. Cresta tests Conductor on four core build tasks: creating new agents, writing test cases, modifying existing agents, and performing root cause analysis. Scoring covers outcome quality, whether the execution path was appropriate, result quality against curated references, and resource efficiency. That is an eval suite for the agent-building process itself, not only for the agents it produces. Engineering lead Renjie Li says the SDK gave the team a way to use the same agentic development harness programmatically.

For companies building agents in-house, the takeaway is where to put the effort: a general-purpose harness underneath, your domain expertise on top, and a scored evaluation loop around both. The halved deployment time is Cresta’s own early-implementation figure, not an independent benchmark. The architecture is one any team can ship against.