Harness, the CI/CD platform company, announced Agent DLC, a product suite that extends its enterprise software delivery infrastructure to govern the full AI agent development lifecycle. The announcement, reported by ITDigest, targets the gap between sandbox prototyping and production deployment that has stalled enterprise agent adoption.
According to internal benchmarks cited by Harness, only 8% of organizations have successfully deployed agentic AI into active production. The company attributes the bottleneck to three concerns: unpredictability of LLM-driven execution paths, data exposure risk, and unmanaged agent behavior in production environments.
Five Products, One Pipeline
Agent DLC consists of five capabilities integrated directly into existing Harness developer pipelines, according to ITDigest:
AI Evals lets developers build evaluation datasets, scoring functions, and automated quality gates that measure agent performance on correctness, safety, and latency before code merges.
Agent Deployments extends enterprise canary releases, manual approvals, and Open Policy Agent (OPA) controls to managed agent environments including Amazon Bedrock AgentCore and Google’s Agent Runtime. The intent is to eliminate cloud-specific deployment silos by applying uniform release governance across providers.
AI Configs decouples prompt engineering and model selection from application redeployments. Teams can test, update, or roll back runtime configurations without rebuilding underlying codebases.
AI Asset Catalog scans code repositories to index every active agent, skill, and plugin while assigning ownership, a measure aimed at preventing what Harness describes as “shadow AI sprawl.”
AgentTrace captures step-by-step execution paths, tool calls, and model decisions across multi-turn sessions for debugging and compliance audits.
Security Across the Agent Lifecycle
The announcement includes security tooling at both pre-deployment and runtime phases. Pre-deployment checks include what Harness calls “Primitive Scanning” to catch skill misconfigurations, automated generation of an AI Bill of Materials (AIBOM), and adversarial testing against the OWASP Top 10 for LLMs. In production, Agent DLC operates as a runtime firewall to block prompt injection attacks and halt data exfiltration attempts.
Harness is also open-sourcing foundational components, including harness-sdk and harness-evals, to let developers integrate tracing and evaluation primitives into custom AI applications.
Building on Recent Launches
Agent DLC extends a broader push Harness has made into the AI development infrastructure market. On June 30, Harness CEO Jyoti Bansal announced Autonomous Worker Agents, AI agents that execute as pipeline steps inside Harness CI/CD with scoped credentials, OPA policy enforcement, approval gates, and full audit trails. Those agents run on customer-controlled infrastructure via Harness Delegates, lightweight runtimes installed inside the customer’s Kubernetes cluster or VPC.
Earlier, in June, Harness launched AI DLC Insights and Cloud & AI Cost Management. As SD Times reported, 94% of engineering leaders surveyed for Harness’s State of Engineering Excellence report said they could not find cost metrics in their current measurement frameworks. Trevor Stuart, senior VP at Harness, told SD Times: “There’s a lot of chatter right now around rising AI costs. But the real challenge isn’t the spend itself, it’s that teams can’t pin down the impact of the spend.”
The Competitive Landscape
Harness is positioning Agent DLC as the CI/CD governance layer for agents in a market where cloud providers, DevOps vendors, and startups are all building overlapping pieces. AWS launched Bedrock AgentCore for agent hosting and runtime management. Google ships Agent Runtime with its Vertex AI platform. IBM’s Watsonx Orchestrate released an Agentic Control Plane in June for centralized agent governance. Box launched agent security controls including prompt injection detection and human-in-the-loop oversight on July 21.
The differentiator Harness is claiming: treating agents as managed workloads within existing software delivery pipelines rather than as standalone AI products requiring separate governance stacks.