Only 8% of organizations have agentic AI in production. That number, cited by Gartner and repeated in Harness’s announcement this week, says more about the state of developer tooling than it does about model quality. The models are fine. Claude Code, Codex, and Jules are all producing solid output. What’s broken is the stack around the model — the testing, the shipping, the monitoring, the incident response. On July 21, Harness announced Agent DLC, an extension of its CI/CD platform that treats AI agents like software, with all the pipeline discipline that implies.
The Gap No One Is Talking About #
A March 2026 survey found 78% of enterprises have active AI agent pilots. Under 15% of those agents have reached production. The industry failure rate converges around 88%. Gartner predicts 40% or more of agentic AI projects will be canceled outright by 2027 — not deprecated, canceled.
The reason is not model capability. It’s the five gaps that account for 89% of production scaling failures: no monitoring tooling, integration complexity, inconsistent quality at volume, unclear ownership, and insufficient domain training. Most of those are solved by applying standard software delivery discipline. Agents have been exempt from that discipline because developers have been shipping them like scripts — fast, informal, and outside the normal release process.
What Harness Built #
Agent DLC extends the Harness platform across five areas. The two most consequential for developers are AI Evals and AgentTrace.
AI Evals
Evals are how you enforce quality gates on non-deterministic systems. You define a dataset of representative inputs, configure scoring functions (correctness, safety, performance — each normalized 0.0–1.0), and set a pass threshold. Harness wires this into your pipeline as a stage, the same way a lint check or test suite runs. If the eval score drops below threshold after a model change or prompt update, the deploy stops. This is the missing link between “we tested it manually” and “it regressed silently in production.”
AgentTrace
AgentTrace records the full decision path for every agent run — which tools were called, latency at each step, how model or prompt changes shifted behavior across a session. The practically useful feature: production failures can be one-click exported into eval cases, closing the loop between what broke in production and what gets tested in CI. Harness is open-sourcing the foundational components (harness-sdk and harness-evals) under Apache 2.0, making them usable with any OpenTelemetry backend without requiring the Harness platform.
AI Configs and Agent Deployments
AI Configs decouples model and prompt changes from agent binary changes. You can swap the underlying model or tune a prompt at runtime without a full redeploy — backed by Harness’s existing feature flag infrastructure. Agent Deployments adds out-of-the-box pipeline steps for Amazon Bedrock AgentCore and Google’s Agent Runtime, including canary releases, OPA guardrails, and approval gates. An AI Asset Catalog rounds out the platform by auto-discovering agents, skills, and plugins across repositories with ownership metadata attached.
Why Traditional CD Breaks for Agents #
Standard continuous delivery assumes deterministic software: the same input produces the same output. An agent is structurally different. The same prompt with the same tools will produce different outputs across runs. That makes traditional canary analysis meaningless — you cannot diff an agent’s behavior the way you diff code. Evals are the substitute. They don’t check outputs; they measure quality across a distribution of outputs. Harness is one of the first major CI/CD vendors to build this directly into the delivery pipeline rather than treating it as a separate ML concern.
The Mindset Problem Is Bigger Than the Tooling Problem #
Harness is correct that agents need the same deployment discipline as regular software. But the reason most teams don’t have that discipline is not that the tooling was missing. It’s that they never admitted their agent was production software in the first place. Pilots become production quietly, without the infrastructure investment, governance docs, or ownership structure that would accompany a normal service release. Harness Agent DLC gives teams a forcing function: when your agent goes through the same change record and approval chain as the rest of your software, it becomes harder to pretend it’s still experimental.
What to Do Now #
- If you’re already using Harness: Agent DLC features are rolling out to customers now. Check the official announcementfor availability and onboarding details. - If you’re not on Harness: the harness-evals libraryis open-source under Apache 2.0 and works with any OTel backend. Start with evals independently of the platform. - Audit your current agent deploy process: do you have eval gates, observability into the decision path, and runtime control over prompts? If not, identify the single gap most likely to catch failures first and close it before the next incident. For most teams, that gap is observability.