AI Agents Get the DevOps Treatment as Governance Moves Into the Toolchain

For two years, most enterprise AI agents were built in notebooks, chat interfaces and no-code builders, outside the pipelines that govern every other piece of production software. This week, a cluster of announcements from Harness, GitLab, AvePoint, Teradata, Axonius and others suggested that era is ending. Vendors are now treating agents like code and like assets: something to version, test, deploy with approvals, inventory and audit. For CIOs, that shift could do more to move agents from demo to production than any new model.

Why are AI agents being pulled into the software delivery pipeline?

Because the old pipelines already solve the problems agents keep creating. Harness, a software delivery platform, launched what it calls the AI Agent Development Lifecycle on July 22, extending the same pipelines and controls that customers use for application code to AI agents. The release includes evaluation datasets with automated quality gates that catch regressions when a model or agent changes, prompt and model configurations managed through feature flags so teams can compare and roll back without redeploying, canary releases to managed agent runtimes on AWS and Google Cloud, and a catalog that discovers agents, plugins and skills across repositories and assigns owners.

Harness also added a trace tool that records agent decisions and model interactions, plus runtime controls aimed at prompt injection, unauthorized tool use and data exfiltration. Harness CEO Jyoti Bansal framed the change as a widening of what counts as software, which he said now includes "agentic code, written across engineering, product, sales, and support teams alike."

GitLab moved in the same direction two days earlier. GitLab 19.2, released July 20, adds agents that open merge requests to fix vulnerable dependencies and a security review flow that looks for logic and authorization flaws that pattern-based scanners miss. The governance details are the point: agent-made changes pass through existing approval gates with a full audit trail, the security review agent never approves its own work, a new report logs AI-assisted actions as audit events, and administrators get access controls over which agents can run through MCP and what they can reach.

GitLab's chief product and marketing officer, Manav Khurana, explained why. Coding agents, he said, "moved the bottleneck downstream to reviews and security."

What problem are vendors trying to solve?

Visibility, mostly. Enterprises adopted agents through many platforms at once, and few can say how many they have or what those agents can touch.

AvePoint, which sells data protection and governance software, cited its own annual State of AI Report finding that 88.4% of organizations had at least one AI agent-related security incident in the past year, in a July 22 announcement. Its new AgentPulse tools back up and restore Microsoft Copilot Studio agents, including prompts and workflows, give visibility into Salesforce Agentforce agents, and apply ownership, sensitivity labels and policy guardrails when agents are discovered. AvePoint CTO John Peluso summed up the pitch: "A trust layer only works if it grows as fast as adoption does."

Elsewhere in the week, Arctera's State of AI Governance 2026 report found that fewer than one in five AI-using organizations can prove their governance controls work, and startup Neo emerged from stealth with $100 million to secure AI software and agents that select tools and act on their own, according to Solutions Review's weekly roundup.

Practitioners at AI Connect events have been describing this gap for months. "You can't govern what you can't see," Snowflake's Arun Agarwal, Senior Product Marketing Manager for AI/ML, said in a AI Connect webinar on moving AI from prototype to production. He added that an agent can reach a correct-looking answer by a flawed path, which is why trace and evaluation tools now sit at the center of these releases.

How do open standards and data access fit in?

Pipelines govern how agents are built and shipped. Standards and data layers govern what agents can reach once they run.

Teradata said on July 22 that it joined the Agentic AI Foundation, the Linux Foundation-hosted group formed in December 2025 that oversees projects including the Model Context Protocol, AGENTS.md, Goose and Agent Gateway. Teradata said it will contribute to MCP's evolution with a focus on regulated, hybrid, sovereign and air-gapped environments, and that its MCP server gives agents governed data access while preserving existing identity and permission controls.

That design echoes a demo Teradata gave at a AI Connect webinar in April, where developer advocate Janeth Graziani built an agent that worked only through approved tools. In that session on building AI agents with MCP tools, she put it plainly: "The agent doesn't generate SQL."

Asset inventory vendors are joining in from the security side. Axonius on July 22 added an AI agent and an MCP server that let AI applications query its asset data in plain language. Its chief product officer, Moshe Ben Simon, argued that every AI tool security teams deploy "is only as good as the asset data behind it."

Does reuse change the economics of agents?

It might. Squirro, an enterprise AI software company, made a catalog of 13 prebuilt agents generally available on July 21, built so that the first deployment sets up connectors, security approvals and a knowledge layer that later agents reuse, the company said. Its CEO, Dave Clarke, said the value lies in "what the second one inherits from the first."

The same logic applies to governance. Once evaluation gates, audit logs and access policies exist for one agent, the second agent should cost less to approve. That is the argument AI Connect panelists made for moving from human in the loop to agent ops, where Revefi's Girish Bhat predicted that data, security and AI governance would converge because "fundamentally it's all about provenance."

What should enterprise leaders take from the week?

The direction is clear even if the tools are young. Several of the features announced are in beta, preview or early access, and vendor statistics come from vendor-sponsored research. Buyers should test claims against their own environments.

Still, the practical checklist is getting sharper:

• Register every agent with a named owner, in the same catalog or CMDB used for other software assets.
• Put agents through the release pipeline, with evaluation gates, versioned prompts and rollback.
• Log agent actions as audit events, not just chat transcripts.
• Route data access through governed interfaces such as approved MCP tools, rather than open-ended queries.

The companies that already run disciplined DevOps have a head start. For everyone else, the lesson of this week is that agent governance is becoming an extension of software governance, not a separate discipline to invent from scratch.

Frequently asked questions

What is an AI agent development lifecycle?

It is the practice of building, testing, deploying, securing and governing AI agents with the same pipelines and controls used for application software, including evaluation gates, versioned prompts, approvals and rollback.

What is the Agentic AI Foundation?

A Linux Foundation-hosted group formed in December 2025 that oversees open agentic AI projects including the Model Context Protocol, AGENTS.md, Goose and Agent Gateway.

How can enterprises govern AI agents today?

Register every agent with an owner, run agents through the release pipeline with evaluation gates, log agent actions as audit events and route data access through governed interfaces such as approved MCP tools.

Filed under: Agentic AI, AI Governance & Trust, Observability & Data Quality