Harness adds CI/CD controls for deploying AI agents

Harness has launched Agent DLC, a product for building, testing, deploying and governing AI agents through the same CI/CD workflows developers use for conventional software. The company did not disclose pricing, customer names or revenue tied to the launch, but it is positioning the product as infrastructure for enterprises that want agents in production without standing up a separate toolchain.

The pitch is practical rather than novel. Harness said Agent DLC combines evaluation, release management, runtime controls, security scanning and observability for AI agents in one system. Those functions already exist in pieces across the market, but Harness argues that enterprises are losing time stitching together evaluators, firewalls, scanners and tracing tools as agent projects move beyond demos.

Harness cited its own internal data showing that 8% of organizations have put agentic AI into production. The company attributes the gap to the difficulty of operationalizing agents safely, because agents can choose different APIs, tools and execution steps each time they respond to a prompt. That variability makes the standard test-once assumptions of deterministic software weaker for agent workloads.

What Agent DLC includes

Agent DLC adds several components to Harness’ software delivery platform. Harness AI Evals lets teams create evaluation datasets and quality gates that run when an agent or model changes, with the stated goal of catching regressions before release. Harness Agent Deployments supports managed agent runtimes and integrates with third-party continuous delivery platforms, including Amazon Bedrock AgentCore, according to the company.

For runtime management, AI Configs handles release changes, prompt management and model updates without requiring a redeployment of older versions. Harness said this enables faster rollback when a prompt or model change creates a problem in production. The company also added an asset catalog inside the Harness Internal Developer Portal to discover and register agents, skills and plugins, an attempt to limit duplicate work and unmanaged tools across teams.

Security is a separate module in the package. Harness said Agent Security scans the models and skills used by agents for misconfigurations, creates a bill of materials for each agent and tests for adversarial inputs before deployment. In production, the company said the module acts as a firewall to enforce policies aimed at blocking prompt injection attacks and data exfiltration.

The observability piece is Harness AgentTrace. It records the actions agents take at the run level and across sessions, so teams can audit how an outcome was produced rather than only inspect the final response. Harness said the telemetry can be used to compare output quality and find performance bottlenecks. The company is releasing foundational components of AgentTrace under an open-source license, though it did not specify adoption targets.

Why it matters for dev platform vendors

The launch shows how software delivery vendors are trying to absorb agent governance into existing developer platforms instead of treating AI agents as a separate operational category. That is a defensive and offensive move for Harness: if enterprises keep agent release processes inside the CI/CD stack, platform vendors can retain budget and workflow control as AI application development changes.

Harness’ claim is that enterprises do not need to replace their release processes to ship agents. The harder test will be whether its guardrails reduce production risk in real deployments, especially for agents that call enterprise systems and handle sensitive data. The company did not disclose benchmarks, incident reduction metrics or deployment volume for Agent DLC.

This story draws on original reporting from SiliconANGLE.