United StatesTechnology and telecom2025 – ongoingDelight.ai
Internal debugging and regression-testing tools for production AI agents, built with Claude Code
Delight.ai cut the time to fix a production AI agent issue from a week to one or two days with internal tools built in Claude Code
Problem
Delight.ai runs AI customer support agents for its clients. Finding why a live agent behaved wrongly, and checking a fix before redeploying it, took the AI/ML team about a week.
Approach
The AI/ML team started using Claude Code in November 2025 to build its own tooling. One tool is a conversation debugger that compares the behavior an agent was expected to show with what it actually did. Another generates and runs simulated conversations at scale to regression-test agents before they reach customers. Claude runs on Amazon Bedrock and through the direct Anthropic API.
Results
Takeaway
Delight.ai's team built its own debugging and testing tools instead of waiting for a vendor to supply them. The company did not say how it measured the drop in fix time.
Categories
- Industry
- Technology and telecom
- Business function
- IT and R&D
- Innovation type
- Process innovation
- AI technology
- Generative AI
- Adoption
- Off-the-shelf tool or platform
- Evidence
- Vendor case study
Sources
Source summaries
Anthropic's customer story describes how Delight.ai's AI/ML team used Claude Code to build a conversation debugger and a simulated-conversation regression testing tool for its production AI agents. It reports that the time to fix and redeploy an agent issue fell from 1 week to 1-2 days. The page carries no date; the Internet Archive first saved it on June 24, 2026.
Full details are in the original sources.