Harness launches tools to manage AI agents lifecycle

Harness has launched Agent Development Lifecycle tools on its software delivery platform, extending its system for managing application code to AI agents.

The release adds products for testing, deployment, configuration, governance and monitoring as companies look for ways to move agent-based projects beyond the pilot stage.

Called Agent DLC, the offering is designed to let enterprises build, test, deploy, secure and operate AI agents through the same delivery pipelines and control processes they already use for software releases.

It includes AI Evals, Agent Deployments, AI Configs, AI Asset Catalog, AgentTrace and a set of security functions covering the full lifecycle of agent-based systems.

Harness cited Gartner data showing that only 8% of organisations have agentic AI in production as evidence that many businesses are still struggling to operationalise the technology.

That gap reflects a basic problem with how AI agents behave. Traditional software is deterministic, meaning the same test on the same code should produce the same result. Agents can produce different outcomes from the same input because underlying language models may choose different tools or actions from one run to the next.

As a result, software testing and incident management methods used for conventional applications do not transfer neatly to agents. Businesses also face added risks when agents interact with tools, APIs and other models, particularly when customer data, compliance rules or operational approvals are involved.

Harness is positioning the launch around that governance problem. Rather than introducing a separate workflow for agents, it says enterprises can manage them through existing pipelines, approvals and policy controls.

New products

AI Evals is intended to measure agent quality by letting teams define evaluation datasets, connect scoring functions and set quality gates to catch regressions when an agent or model changes.

Agent Deployments extends release management controls already used in Kubernetes environments to managed agent runtimes, including Amazon Bedrock AgentCore and Google's Agent Runtime. This lets agents move through the same deployment processes as other software components.

AI Configs focuses on prompts and model changes at runtime. Teams can test prompt variants, assess results and roll back changes without redeploying an agent.

AI Asset Catalog is designed to discover agents, skills and plugins across repositories and link each one to an owner. The aim is to give organisations a clearer inventory of what is being built and what is running.

AgentTrace records events during a single agent run and across a wider session, showing the path taken by an agent and where delays occur. Harness is also open-sourcing the components behind AgentTrace, including harness-sdk and harness-evals, so developers can use the tracing tools in their own AI applications.

Several existing products will also apply to agents without modification, including Continuous Integration, Artifact Registry, AI Test Automation and AI Cost Management.

Security layer

Security is a central part of the launch. Static scanning tools do not address the risks posed by agents that can choose their own actions, connect to external systems and create sub-agents.

The new security functions include Primitive Scanning, which checks for misconfigurations in skills, prompts and models, and AIBOM, which records the models, tools and dependencies used to build an agent. AI Testing is designed to run agents against adversarial inputs and the OWASP Top 10 for large language models.

On the operational side, Agent Discovery and Posture Management is intended to surface active agents, map how they interact and assess their posture across an organisation. AI Firewall is designed to enforce policy in real time against prompt injection, tool misuse and data exfiltration.

According to Harness, these features create a single audit trail from development to production. The launch also ties into its Software Delivery Knowledge Graph, which connects data from different stages of the software delivery process across both applications and agents.

Recently, the company introduced Autonomous Worker Agents for use inside software delivery pipelines.

"When we started Harness, the vision was a safety harness for code," said Jyoti Bansal, Co-Founder and Chief Executive Officer, Harness. "Until recently, that meant application code. Today it also means agentic code, written across engineering, product, sales, and support teams alike, each building agents for their own workflows. Everything you've done for software delivery over the last decade - governance, orchestration, security, testing - you can now do for agents in the same platform."

Agent DLC is rolling out now to customers.