Harness Survey Finds AI Agent Controls Lag Enterprise Confidence

Harness detailed The State of Agent DLC 2026 in a press release, based on a survey of 700 engineering and technology leaders at large enterprises. Sapio Research conducted the survey in July 2026 across the United States, United Kingdom, France, Germany, and India.

The survey found that 77% of respondents believed they had a complete inventory of agents, MCP servers, and large language models, but 44% used active discovery tools. While 74% were confident their testing could catch a production impacting failure, 19% had an automatic gate that blocks every bad release. Another 76% said they could disable a misbehaving agent within 15 minutes, while 33% had an instant kill switch.

Among respondents who considered their agents secure from end to end, 88% reported security incidents. The overall incident rate was 87%. Although 74% said they understood spending per agent, 60% exceeded their budget in the previous quarter.

Only 53% of changes involving agents passed through a standard pipeline before production. The survey also found that 58% of organizations recorded more production incidents per 100 changes after deploying AI agents.

Harness sells the delivery and governance tooling that the gaps in this survey describe, so the findings run in the direction of its own product. The internal contradictions are the more durable part: the same respondents who claim complete inventories mostly do not run discovery, and the ones who call their agents secure report incidents anyway. Those comparisons hold whatever the survey was designed to show.