SAN FRANCISCO – Harness, the AI Software program Supply Platform firm, immediately introduced it’s extending its platform to cowl the complete AI Agent Growth Lifecycle (DLC), giving enterprises a single set of pipelines and controls to construct, take a look at, deploy, and run brokers the identical manner they already ship the whole lot else.
Each enterprise is constructing AI brokers, however most can’t get them previous inside pilots or proofs of idea. In response to Gartner®, “Solely 8% of organizations have agentic AI in manufacturing.” The software program supply lifecycle enterprises’ belief for delivery utility code hasn’t prolonged to brokers but, trapping the ROI of inside AI investments. Actual innovation arrives as soon as an organization can run an agent reside with the identical belief and confidence it has in the remainder of its software program.
“After we began Harness, the imaginative and prescient was a security harness for code,” stated Jyoti Bansal, co-founder and CEO of Harness. “Till lately, that meant utility code. At present it additionally means agentic code, written throughout engineering, product, gross sales, and help groups alike, every constructing brokers for their very own workflows. The whole lot you’ve completed for software program supply over the past decade — governance, orchestration, safety, testing — now you can do for brokers in the identical platform.”
Why AI brokers break the normal software program supply lifecycle
Conventional software program works as a result of it’s predictable. Software code is deterministic. Run the identical take a look at towards the identical code twice, and it produces the identical outcome each occasions.
Brokers don’t work that manner: an agent’s underlying language mannequin decides the best way to full a job, and the identical agent, given the identical enter, can select a unique device or take a unique motion from one run to the subsequent. A take a look at that passes as soon as gives no assure it’s going to move the subsequent time. Incidents cease being reproducible on demand, which implies the usual playbook for catching and fixing bugs doesn’t switch both.
The stakes rise with the scale of the enterprise. A rogue agent can expose buyer information, violate a compliance coverage, or take an motion no person authorised. Enterprises want a option to reply for what their brokers are doing, and the normal software program supply lifecycle was by no means constructed to provide them one.
New Harness Agent DLC merchandise and capabilities
Agent DLC closes the hole between constructing an agent and delivering it safely to manufacturing. At present’s launch contains 5 new merchandise and capabilities spanning testing, deployment, operations, and governance:
Harness AI Evals makes agent high quality measurable, letting groups outline eval datasets, wire up scoring features, and set high quality gates that routinely catch regressions at any time when an agent or mannequin modifications.
Agent Deployments lengthen the canary releases, approvals, and OPA guardrails that Harness already applies to Kubernetes deployments to managed agent runtimes like Amazon Bedrock AgentCore and Google’s Agent Runtime. Brokers now ship by way of current pipelines as an alternative of a separate cloud-specific workflow.
AI Configs help the discharge and administration of prompts and mannequin modifications at runtime, backed by the identical characteristic flagging infrastructure that already manages code releases. Groups can take a look at what performs finest and roll again immediately, with out redeploying.
AI Asset Catalog routinely discovers each agent, ability, and plugin constructed throughout a corporation’s repositories and hyperlinks every to an proprietor, so nothing ships or runs unaccounted for.
Harness AgentTrace information what occurs throughout a single agent run and throughout a full multi-step session, displaying which path an agent took, the place it slowed down, and the way completely different fashions or prompts have an effect on the result. Harness can also be open-sourcing the foundational elements behind AgentTrace, together with harness-sdk and harness-evals, so builders can deliver the identical tracing primitives into their very own AI functions.
As well as, current Harness merchandise already lengthen to brokers with out requiring any modifications: Steady Integration builds them like every other service, Artifact Registry tracks their variations and dependencies, AI Check Automation validates their responses in plain English standards, and AI Value Administration extends spend visibility to each agent and mannequin.
Securing the Agent DLC
Brokers select their very own strategy and path to get there, so their conduct is difficult to foretell and simply as exhausting to safe. They increase their very own assault floor by connecting to instruments and APIs, spawning sub-agents, and inheriting belief from each mannequin they contact. Static scans had been by no means designed for this sort of threat. Harness is launching new safety capabilities to shut that hole.
Shift-left: constrain what brokers can do earlier than they ship.
Primitive Scanning flags misconfigurations in agent abilities, prompts, and fashions.
AIBOM captures each mannequin, device, and dependency an agent was constructed with.
AI Testing runs brokers towards adversarial inputs and the OWASP Prime 10 LLM and Agentic AI dangers.
Defend-right: implement coverage and keep visibility as soon as they’re reside.
Agent Discovery and Posture Administration constantly maps brokers as they spin up and the way they join throughout the group.
AI Firewall enforces coverage in actual time towards immediate injection, device misuse, and information exfiltration.
Collectively, these capabilities give Agent DLC a single audit path from growth to manufacturing.
Constructed on the Harness platform
Harness constructed context and intelligence straight into the platform with the Software program Supply Information Graph, which captures and connects information from each stage of the supply lifecycle, now spanning each functions and brokers. Organizations counting on siloed instruments don’t have that very same related view.
In June 2026, Harness launched Autonomous Employee Brokers, a platform for constructing and safely working AI brokers inside software program supply pipelines. Employee Brokers run as ruled steps inside these pipelines, lined by the identical controls Harness already applies to each deployment.
Agent DLC extends that very same context and governance throughout the complete agent lifecycle. The pipelines, insurance policies, approvals, and proof that already apply to a corporation’s code now apply to its brokers too, so eval gates, deployment approvals, and safety checks run as phases inside a single pipeline, from the second an agent is created by way of the whole lot it does afterward.
Availability
Harness Agent DLC capabilities are rolling out now to Harness clients. For a full breakdown of what’s included at every stage of the lifecycle, go to [blog page].
How does Harness AI Evals work for testing AI brokers?
Harness AI Evals lets groups outline analysis datasets, configure scoring features, and set high quality gates that routinely detect regressions when an agent or underlying mannequin modifications. It addresses the non-deterministic nature of LLM-based brokers, the place the identical enter can produce completely different outputs throughout runs.
Can Harness Agent DLC deploy brokers to Amazon Bedrock AgentCore and Google Agent Runtime?
Sure. The Agent Deployments functionality extends Harness canary releases, approval workflows, and OPA coverage guardrails to managed agent runtimes together with Amazon Bedrock AgentCore and Google’s Agent Runtime, so brokers ship by way of current Harness pipelines somewhat than cloud-provider-specific workflows.
How does Harness safe AI brokers towards immediate injection and LLM vulnerabilities?
Harness introduces each shift-left and shield-right safety capabilities. Shift-left instruments embrace Primitive Scanning for misconfigurations, AIBOM for dependency monitoring, and AI Testing towards OWASP Prime 10 LLM and Agentic AI dangers. Defend-right capabilities embrace Agent Discovery and Posture Administration for runtime visibility and an AI Firewall that enforces coverage towards immediate injection, device misuse, and information exfiltration in actual time.
