Autoheal raises $7.9 million to evaluate and repair enterprise AI agents

Autoheal has raised $7.9 million in seed funding led by Innovation Endeavors to expand a platform that evaluates AI agents used in software operations and proposes changes when performance falls short. It targets enterprises that need to measure and repair multiple agents across production workflows.

The round also included Emergent Ventures, U&I Ventures, Darkmode Ventures, Batch Ventures and Param Hansa Values. U&I Ventures said it invested in Autoheal, corroborating the financing. Autoheal said Harpinder Singh of Innovation Endeavors will join its board in connection with the round. The company did not disclose its valuation or ownership terms.

Autoheal describes its product as a control layer that connects existing coding agents with code repositories, continuous-integration and deployment systems, observability tools, cloud runtimes and issue trackers. Those systems feed a shared engineering context graph that retains information about how agents work and what happens after they act.

According to Autoheal’s company-issued financing announcement, an Evaluator agent scores other agents’ work using downstream signals. A separate Healer agent can then propose pull requests that alter the agents’ skills, prompts, tools or model selections. Autoheal says those proposed behavior changes are version-controlled, tested against historical benchmarks for regressions and require engineer approval.

Autoheal also plans to use private enterprise engineering data to train smaller, customer-specific models that operate inside customer-controlled environments. The company did not provide a detailed allocation for the new capital beyond scaling the platform and pursuing that product roadmap.

Autoheal and customers quoted in its announcement said the system reduced some investigation or root-cause timelines from hours to minutes and freed engineering capacity. Those performance figures remain company and customer claims: the announcement did not disclose a measurement methodology, and no independent audit of the results was identified.